# Jared Smith — Founder & Engineering Leader > Founder and engineering leader (13+ years) who builds and runs the security and engineering function — at his own companies and fractionally for pre-Series-A AI startups and scaling teams. Production AI, distributed backend systems, resilient multi-cloud infrastructure. ## About Jared Smith is a founder and engineering leader based in California. He co-founded PopSocial and scaled engineering from 1 to 15+, then led security as CISO at Lavender while writing production code, and shipped as a senior engineer at BlockFi, InsideTrack, and AAMP Global — pairing hands-on delivery with architecture that holds up under production load. Specialties: production LLM systems, agentic AI workflows, distributed systems and observability, cloud cost optimization across GCP / AWS / Azure, SOC 2 / HIPAA / ISO 27001 readiness, and engineering leadership at the small-team-to-fifteen scale. Writing voice: pragmatic, opinionated, senior — first-person operator essays grounded in real numbers from real roles (Lavender, BlockFi, InsideTrack, AAMP Global, PopSocial). No abstract theorizing, no listicles, no "in theory you could." Every claim ties back to a shipped outcome. Contact: jared@sublimecoding.com — fractional engagement details at https://sublimecoding.com/consulting A more detailed version of this file with embedded content bodies is available at https://sublimecoding.com/llms-full.txt — the full text of every current essay (flagship and recent alike), concatenated into one ingestion-ready document. ## Key pages - [Home](https://sublimecoding.com/): Overview of services, credentials, and engagement model - [About](https://sublimecoding.com/about): Positioning page — who Jared helps, how engagements work, and the operator background behind the practice - [Resume](https://sublimecoding.com/resume): Full work history with named outcomes and selected metrics - [Start Here](https://sublimecoding.com/start): Curated reading paths for founders/CTOs, VPs of Engineering, and senior engineers - [Writing](https://sublimecoding.com/blog): Long-form essays on AI engineering, security, and engineering leadership - [Topics](https://sublimecoding.com/blog/topics): Complete tag directory across the writing — every topic, post counts, ordered by depth of coverage - [Consulting](https://sublimecoding.com/consulting): Eight engagements with prices: Security Review Sprint ($3,000 fixed, one security questionnaire up to 250 questions in 5 business days), Fractional Chief AI Officer (AI and coding-agent adoption for engineering teams) and Fractional CTO at $3,500/month, Advisory at $3,500/month, Security Leadership (vCISO) at $2,500/month, plus scoped AI Implementation, Embedded Engineering and Technical Assessment - [Playbook](https://sublimecoding.com/playbook): The 90-Day AI Startup Security Playbook — eight flagship essays compiled into one downloadable PDF + browsable HTML - [AI Vault](https://sublimecoding.com/vault): The Senior Engineer's AI Vault: an agent-first setup kit where Claude Code or Codex builds a memory vault, hooks, handoffs and a verify script, plus 50 prompts. Pay what you want from $14.99 - [FAQ](https://sublimecoding.com/faq): Direct answers to the questions founders most often ask before booking - [Products](https://sublimecoding.com/products): Paid kits your coding agent installs (Coding Agent SOC 2 Kit, The Distribution Stack, AI Vault) and the free tools and Security Playbook, in one grid - [Tools](https://sublimecoding.com/tools): Interactive tools for founders and engineering leaders — including the vCISO cost calculator that sizes security-leadership spend by stage and team - [Privacy](https://sublimecoding.com/privacy): Cookie, analytics, and data-handling policy - [Sitemap](https://sublimecoding.com/sitemap.xml): All indexable URLs ## Topic guides - [AI engineering](https://sublimecoding.com/topics/ai-engineering): What production AI engineering actually looks like in 2026: the autonomy ladder for agents, the workflow shift, telemetry, and team sizing. - [Security for startups](https://sublimecoding.com/topics/security-for-startups): A guided reading order for startup security: what to read first on SOC 2 as a revenue tool, vCISO hiring, and securing AI-native products. - [Engineering leadership](https://sublimecoding.com/topics/engineering-leadership): Engineering leadership at startup scale: hiring from one engineer to fifteen, rituals that work at small teams, the staff-engineer interview loop. - [Elixir and the BEAM for AI systems](https://sublimecoding.com/topics/elixir-for-ai): Why language choice matters for AI systems: BEAM concurrency for agents, what Go frameworks cost you, and why Elixir is the language AI writes best. - [Omarchy](https://sublimecoding.com/topics/omarchy): Running Omarchy with coding agents, starting on an M1 MacBook: the install, what breaks, the keyboard, and what a security lead inherits from agents. ## Selected writing — flagship essays - [An AI Just Deleted a Production Database in Nine Seconds. Hire More Engineers.](https://sublimecoding.com/blog/ai-deleted-production-database-hire-more-engineers): Replit's AI agent ignored a code freeze and wiped a production database in nine seconds. The strongest case yet for hiring more senior engineers in the AI boom, with the supervisory architecture that prevents the next incident - [Amazon Let the AI Drive. It Hit a Tree.](https://sublimecoding.com/blog/amazon-let-the-ai-drive): Amazon mandated AI coding, then let its Kiro tool change production infrastructure unsupervised — a 13-hour outage and millions of lost orders followed. The remediation, in Amazon's own handwriting, was two-person review per deploy: the case for more skilled humans in the loop, not fewer, written by the company that bet the other way - [AI Won't Shrink Your Team — It'll Expose Why You Needed a Bigger One](https://sublimecoding.com/blog/ai-wont-shrink-your-team): Why AI multipliers don't replace headcount, and what they reveal about the hiring decision you avoided - [Nadella Is Right About AI and the Firm. Mostly.](https://sublimecoding.com/blog/nadella-token-capital-small-teams): A founder's response to Satya Nadella's 'human capital and token capital' essay — agreeing AI amplifies judgment, but translating the enterprise learning loop down to small-team scale (CLAUDE.md as the RL environment, the review gate as evals) and pricing the platform-serving prescription - [Agents Don't Need Prompts. They Need Managers.](https://sublimecoding.com/blog/agents-dont-need-prompts-they-need-managers): Agents stop before the work is done, and the useful read is not that the model is lazy. A person arrives wrapped in a management system — queue, standup, a manager who notices the silence, a peer who catches the miss — that nobody had to build for them. An agent arrives with none of it. What each of those functions looks like as something you actually build: state on disk rather than in the conversation, gates on output rather than effort, and a named human whose job includes noticing the stop - [Nobody Got Replaced. Agents Got Added.](https://sublimecoding.com/blog/nobody-got-replaced-agents-got-added): The viral claim that McKinsey swapped 25,000 employees for AI agents is false — the transcript says 60,000 total, about 40,000 humans plus agents stacked on top, with client-facing roles up 25% while non-client-facing roles shrank. Why fabricated displacement numbers travel as capital-allocation instruments, and the question that replaces 'how many can we cut': what human-to-agent supervision ratio can you actually defend - [How to Actually Make Money With AI in 2026: An Engineer Ranks 13 Side Hustles](https://sublimecoding.com/blog/ai-side-hustles-ranked-by-an-engineer): Thirteen ways people are told to make money with AI in 2026, ranked S through D by an engineer who builds the systems rather than sells the course — graded on upside ceiling, moat, technical leverage, time to first dollar, and fulfillment burden. Website generation is D because the tools compressed the margin to an afternoon of Claude Code. Voice agents keep a real moat and still rank C on a latency budget where ElevenLabs’ ~75ms of synthesis is the easy part. The automation agency is the highest-confidence A, and the two constraints nobody selling it mentions are n8n’s Sustainable Use License and a token COGS of $1,800 to $3,500 per developer per month against a $200 seat. S tier is distribution, which is not an idea at all - [I Graded My Blog Against a 2014 Marketing Course](https://sublimecoding.com/blog/graded-my-blog-against-2014-marketing-course): A founder grades his own site against John Sonmez's 2014 developer-marketing course. Production passes: schema, FAQ blocks, llms.txt, topic hubs, two to three posts a week. Distribution fails: 63 search clicks on 5,612 impressions in 92 days, 126 of 145 pages with zero clicks, first-page money pages earning nothing, three LinkedIn posts against about 70 published, and no email list. What the 2014 advice still gets right, what platform changes and AI answer engines made obsolete, and the changes planned as a result - [AI Made Tokens Cheap. It's Making Hardware Costly.](https://sublimecoding.com/blog/ai-tax-hardware-costs): Apple's June 2026 price hikes across Mac, iPad, and HomePod, read as a signal: the same AI demand that drives per-token inference cost down is bidding the physical substrate — memory, HBM, wafers — up. Why intelligence keeps getting cheaper while the hardware to run it doesn't, and what that two-sided cost curve means for anyone building on AI - [When to Trust an Agent and When to Step In](https://sublimecoding.com/blog/when-to-trust-an-agent-and-when-to-step-in): Four-level autonomy ladder for agentic AI systems and five failure signals that demand human takeover - [How I Triage a New Codebase in 90 Minutes](https://sublimecoding.com/blog/triage-a-new-codebase-90-minutes): Step-by-step protocol for rapid orientation in unfamiliar code - [My Daily Agentic AI Workflow](https://sublimecoding.com/blog/my-daily-agentic-ai-workflow): The actual stack — Claude Code, OpenAI Codex — and the loop that makes them ship rather than break things - [4,154 Commits in Six Months With AI Agents](https://sublimecoding.com/blog/4154-commits-six-months-ai-agents): 4,154 commits in one repo in six months — what changed when agents wrote the code, what got better (velocity, ticket hygiene, language-agnostic shipping), and what got worse (blast radius, regression surface, the craft feeling) - [The One File Where Your AI Model Names Live](https://sublimecoding.com/blog/model-routing-file-ai-coding-agents): Why model names scattered across every repo's config rot the day a model is deprecated or repriced, and the fix: one git-versioned markdown table mapping task type to effort tier and model, imported into every agent session, with a dated note on how the table's own top row moved from Opus 4.8 to Opus 5 - [Your Coding Agent Has No Reason to Write Good Code](https://sublimecoding.com/blog/why-ai-agents-write-unmaintainable-code): Coding models are trained with a binary reward — did the failing tests pass without breaking the passing ones — and maintainability is never graded, because a test verdict costs seconds while bad architecture bills you in quarters. Why the lights-off software factory fails at the model layer rather than the harness layer, what the 2026 Faros telemetry shows about review collapse, and the four places to spend human attention instead - [The Hard Part Was Never the Code. It Was the Theory.](https://sublimecoding.com/blog/code-was-never-the-hard-part-theory-building): Why the phrase "code was never the hard part" reads as obviously true to some engineers and as an insult to others, resolved through Peter Naur's 1985 "Programming as Theory Building": programming produces a theory held by people, and the program text is a lossy, secondary artifact of it. Naur's three tests for whether someone holds a program's theory, his two field reports (group B failing to extend a compiler despite full documentation and the original authors on call; the 200,000-line system whose fault-diagnosis team could not conceive of a document that would help them), why a coding agent samples from the set of changes that are all correct but only some of which conform to the theory, and why an agent-built system can arrive having never been alive in Naur's sense - [Your Junior Dev Leans on the Agent. Good.](https://sublimecoding.com/blog/junior-dev-leaning-on-the-agent): What to do about the junior engineer who builds almost entirely with an agent and produces working code that is rigid, unabstracted, and hard to read. Why telling him to use AI less is advice with an expiry date, why the readability bar was always a proxy for the cost of a human edit, where the quality bar moves instead (tests sourced separately from the code, mutation checks, boundary cases, resilience), the line where structure is still load-bearing, and what a junior genuinely loses when the unglamorous pile disappears - [The 'Smallest Possible Slice' Heuristic for Shipping Complex Features](https://sublimecoding.com/blog/smallest-possible-slice-shipping-complex-features): Better than MVP — how to define a slice that ships in days and earns the next slice - [How to Manage a 4-Person Engineering Team Without Becoming a Manager](https://sublimecoding.com/blog/managing-a-four-person-engineering-team): Rituals, decisions, and the one-on-one cadence that work at small-team scale - [How I'd Hire a Staff Engineer at an AI Startup](https://sublimecoding.com/blog/how-id-hire-a-staff-engineer-at-an-ai-startup): Sourcing, interview loop, and the failure modes that produce mis-leveled hires - [The Perfect Hire Is Killing Your Team](https://sublimecoding.com/blog/perfect-hire-killing-your-team): Why the algorithmic-whiteboard interview measures the wrong thing — recall under pressure and pedigree are among the weakest predictors of job performance (Google's own data; the 2022 selection-validity reanalysis), while 'I'd look it up' is a senior answer in the LLM era. Hire for the slope, not the intercept: effort, ownership, coachability, and the fact that good teammates are forged by the team, not found pre-perfect - [The Pre-Series-A AI Startup Hiring Plan: Who to Hire, in What Order, and Why Most Get It Wrong](https://sublimecoding.com/blog/pre-series-a-ai-startup-hiring-plan): Sequenced guide from first engineer to a fifteen-person team - [The Ruby to Elixir Migration That Cut Our Service Footprint From Ten to Six](https://sublimecoding.com/blog/ruby-to-elixir-migration-ten-to-six-services): Production migration case study — 450K students, 900+ universities, on-call alerts halved - [Your AI Product Needs a Telemetry Layer Before It Needs a Better Model](https://sublimecoding.com/blog/your-ai-product-needs-telemetry-before-better-model): Observability architecture for LLM systems using LangSmith, Helicone, and custom instrumentation - [Surviving Technical Due Diligence as an AI Founder](https://sublimecoding.com/blog/surviving-technical-due-diligence-ai-founder): What a technical due-diligence reviewer opens first — commit history, model provenance, eval reproducibility, LLM lock-in, data-rights trail — and the gaps not worth scrambling to fix pre-Series-A ## Omarchy and DHH, as of October 24, 2026 - [Omarchy Review: Seven Weeks, Three Machines, One Verdict](https://sublimecoding.com/blog/omarchy-review-seven-weeks): Published October 23, 2026. Jared is keeping Omarchy after about seven weeks on three machines (an RTX 3070 desktop, an M1 MacBook Pro and an M2 Pro): Proton Mail works as a web app, backups moved to Restic plus B2, iMessage is still missing, and the 4.5 release candidate drops OBS and Kdenlive from new installs - [I Put Omarchy on an M1 MacBook Pro](https://sublimecoding.com/blog/omarchy-asahi-m1-macbook): Published September 7, 2026, with a status note dated October 21, 2026. The first-week report on installing Omarchy on an M1 MacBook Pro through omarchy-mac, the path that existed before Apple Silicon support was official; the status note adds that Omarchy M is now the official project, the 4.5 release candidate merges Omarchy Mac into Omarchy, and Jared is keeping it - [Linux Setup Was a Hazing Ritual. Agents Changed It.](https://sublimecoding.com/blog/linux-setup-hazing-agents): Published October 4, 2026. Six Linux breakages across an M1 and an x86 desktop, most fixed in five minutes or less with an agent by Jared's estimate, against two to five hours each without one - [I Read Omarchy's AGENTS.md So You Don't Have To](https://sublimecoding.com/blog/omarchy-agents-md-source-read): Published September 11, 2026. Omarchy's AGENTS.md read from the quattro branch: a 133-line router into seven task skills, and five design docs as of 2026-09-07 in which every plan that reached revision 2 names an adversarial Codex review (plans/nix.md was removed upstream on September 9) - [Omarchy Shortcuts on a Mac Keyboard: What Changes](https://sublimecoding.com/blog/omarchy-mac-keyboard-shortcuts): Published September 18, 2026. The Mac keyboard reference: Super is Command on a Mac and the Windows key on a PC, Super + K lists every binding, and about fifteen of Omarchy's 200+ bindings change on Apple hardware - [Omarchy Runs Your Coding Agent With the Safety Off](https://sublimecoding.com/blog/omarchy-agent-safety-off): Published September 16, 2026. The security read of Omarchy's agent launcher: eleven of the fourteen agents it knows start with approvals skipped by default, what Omarchy gets right (mandatory LUKS, a default-deny firewall, a fail-closed sudo timer), and what a vCISO would require before it goes on a company laptop - [Omarchy on a Company Laptop: The SOC 2 Verdict](https://sublimecoding.com/blog/omarchy-company-laptop-soc2): Published September 23, 2026. The SOC 2 verdict on Omarchy as a company laptop: yes, under six controls you add yourself (screen lock, MDM enrollment, patch-cadence evidence, credential consolidation, backup, and agent approval policy), since only LUKS and a default-deny firewall ship ready - [Would I Trust Agents to QA My Daily Driver?](https://sublimecoding.com/blog/agentic-qa-omarchy-trust): Published September 27, 2026. Omarchy's agent-driven release QA boots every release in a fresh VM; Jared trusts it for installer and update regressions, it cannot catch behavior that works as designed but shouldn't, and the diagnose step's permissions are an open question - [DHH's Rails World 2026 Keynote: Pencils Down, Now What](https://sublimecoding.com/blog/dhh-rails-world-2026-keynote): Published September 24, 2026. Commentary on DHH's Rails World 2026 keynote: 37signals going pencils down on hand-written code, the Basecamp 5 story read as an ownership problem, and DHH's Omarchy progress report (install time down from 3 minutes 33 seconds a year earlier to 9 seconds in the lab) - [DHH's Rust Benchmark Measured the Agent, Not Elixir](https://sublimecoding.com/blog/dhh-campfire-benchmark-measured-the-agent): Published October 6, 2026, updated October 24, 2026. DHH's agent-written Campfire ports: Rust got 398 commits in six days against three each for Elixir and Go, and after shared caching and query work a new harness measured the room page at Elixir 5,350, Go 53,060 and Rust 106,494 requests per second ## Elixir, AI coding & the BEAM - [Elixir Is the Language AI Codes Best](https://sublimecoding.com/blog/elixir-ai-codes-best): Benchmark evidence that Elixir tops LLM code-completion rates, and the boring language properties behind it - [Elixir's BEAM Is the Runtime AI Agents Want](https://sublimecoding.com/blog/elixir-beam-ai-agents): Why the BEAM's supervised, isolated processes are the runtime long-lived, crash-prone AI agents actually want - [Elixir's Concurrency Model Is the One You Actually Want](https://sublimecoding.com/blog/elixir-concurrency-model): How the actor model and supervision trees solve failure isolation — not just scheduling like async/await and goroutines - [Why Every AI Agent Framework Is Written in Go (And What That Costs You)](https://sublimecoding.com/blog/go-ai-agent-frameworks): Why agent infrastructure is overwhelmingly written in Go, and what that costs you at the reasoning layer - [Ruby Isn't Dead, It Got Boring — And Boring Is Why It Ships](https://sublimecoding.com/blog/ruby-boring-ships): Why Rails 8's boring, dependency-shedding maturity is a shipping advantage, not a sign of decline - [Elixir Libraries Discord and Supabase Open-Sourced](https://sublimecoding.com/blog/discord-supabase-elixir-libraries): The Elixir libraries Discord and Supabase open-sourced (not Discord-bot libraries), audited for 2026 maintenance status and given a keep/skip verdict for a five-person team. Discord: 13 public repos, 4 active (manifold, sorted_set_nif, instruments, jemalloc_info), 6 abandoned — fastglobal is superseded by :persistent_term (OTP 21.2, same constant-pool zero-copy read and same global-GC-on-write tradeoff), gen_registry by Registry, dispenser by GenStage. Supabase: realtime and supavisor are actively developed but are services to operate, not dependencies; libcluster_postgres and plug_caisson are the small USE picks. Whatnot, Ramp, and PepsiCo publish no Elixir repos; Adobe's elixir-styler is the one library outside Discord/Supabase worth adding today - [15 Elixir Libraries I Reach For in 2026](https://sublimecoding.com/blog/elixir-libraries-i-reach-for-2026): The Hex packages installed on day one of a new Elixir project — including the AI-era additions (Instructor, Req, Oban, rag tooling) — with honest dropped-this-for-that calls - [Build an AI Agent Loop in 50 Lines of Elixir](https://sublimecoding.com/blog/build-ai-agent-loop-50-lines-elixir): The entire AI agent loop pattern — observe, decide, act, repeat — in a single 50-line GenServer module that calls the Anthropic API and dispatches tools, with the three guardrails (max iterations, token budget, receive timeout) that separate a toy from a tool - [I Read Oban's Source Code So You Don't Have To](https://sublimecoding.com/blog/oban-source-code-read): Oban's mechanics read from oban-bg/oban at v2.23.1 rather than the README — job fetching as a CTE wrapping FOR UPDATE SKIP LOCKED specifically to fence the query planner, a leader-elected Stager that polls once a second and notifies queues so workers never poll, and leader election that is one INSERT ... ON CONFLICT whose WHERE clause matches only the inserting node, which is simultaneously why a challenger fails and how the sitting leader renews. The cost arrives on the same line as the reliability: sustained high-UPDATE churn on one table, dead tuples outpacing autovacuum, and SKIP LOCKED still having to walk dead rows, so bloat slows the fetch itself and not just VACUUM - [Oban as a Durable AI Agent Runtime in Elixir](https://sublimecoding.com/blog/oban-durable-ai-agent-runtime): Using plain open-source Oban as a durable AI-agent runtime — idempotent steps, retry-on-429, and a self-enqueuing ReAct loop persisted to Postgres that survives a deploy mid-run (with the honest free-vs-Pro Workflow boundary) - [12 Phoenix LiveView Patterns for Production](https://sublimecoding.com/blog/phoenix-liveview-patterns-production): Twelve battle-tested Phoenix LiveView 1.1 patterns, each paired with the production failure it prevents — streams vs assigns, assign_async, connected? PubSub gating, and the dead-render/connected-socket boundary where most bugs live - [Build an MCP Server in Phoenix With Hermes](https://sublimecoding.com/blog/build-mcp-server-phoenix-hermes): Ship a real MCP server in Phoenix with Hermes — one typed, authorized tool exposed to Claude over Streamable HTTP, with authentication in the server's init/2 reading the transport headers, and the hard rule that you never hand an LLM raw database access - [Streaming LLM Tokens in LiveView, the 2026 Way](https://sublimecoding.com/blog/streaming-llm-tokens-liveview-2026): Token-by-token LLM streaming into Phoenix LiveView without React — start_async plus Req's function-form :into so cancellation is just process death, the OpenAI Responses API, and the backpressure and lifecycle failure modes tutorials skip - [RAG in Phoenix: Hand-Rolled pgvector or Arcana?](https://sublimecoding.com/blog/phoenix-rag-pgvector-arcana): RAG in Phoenix without a separate vector database — the hand-rolled pgvector path (one migration, one Ecto cosine_distance query) versus when the Elixir pipeline libraries Arcana and rag_ex earn their place, and where Postgres is genuinely the wrong call - [Who Actually Runs Elixir in Production, Fact-Checked](https://sublimecoding.com/blog/elixir-in-production-fact-checked): The viral "15 Most Valuable Companies Using Elixir" chart checked row by row against primary sources, with a four-tier evidence scale (first-party engineering content, live job postings as hiring signal only, named third-party reporting, agency listicles as non-evidence) and a currency verdict per row. 9 of 15 hold as CURRENT, 3 are AGING, Pinterest is STALE (the 2025 account of Guardian is Kafka/Flink/Iceberg with no Elixir), Spotify and Together AI are UNSUPPORTED. Whatnot's Elixir Live Service held 583K concurrent viewers in February 2026 and the chart still mislabeled it as the marketplace backend; Adobe's Elixir is Frame.io's; SumUp's payments core is Erlang. Includes a first-person correction of the Together AI post, whose only source was a job posting since removed with no archive - [Why We'd Pick Elixir for an AI Startup Backend](https://sublimecoding.com/blog/why-elixir-for-ai-startup-backend): The case for an Elixir/Phoenix backend at an AI startup — the BEAM's supervised, isolated concurrency as the runtime for agents, token streaming, durable Oban jobs, and pgvector RAG, collapsed into one deployable instead of a fleet of services - [Testing AI Agent Outputs in Elixir with ExUnit](https://sublimecoding.com/blog/testing-ai-agent-outputs-exunit): An ExUnit suite for agent loops in Elixir — asserting on tool-call order and hallucination boundaries rather than final strings, plus evals that survive model upgrades - [TDD With Claude Code in Elixir: What Holds Up](https://sublimecoding.com/blog/tdd-claude-code-elixir): Where TDD with Claude Code actually holds up in an Elixir/Phoenix codebase — ExUnit async, mix precommit, Ecto.Multi idempotency — and where the red-green loop breaks down - [10 Ecto Query Patterns for N+1 and Beyond](https://sublimecoding.com/blog/ecto-query-patterns-n-plus-one): Ten Ecto query patterns organized around the decision layer most N+1 tutorials skip — which shape of data gets which pattern and why. The spine: separate queries (preload) for has_many, a join for belongs_to, because one avoids row-explosion and the other avoids a round trip. Then what you reach for when that rule runs out — filtered preloads, nested trees, composable query modules, Dataloader when you don't control the query shape, subqueries for latest-N-per-group, denormalizing instead of joining, and raw SQL when the DSL is fighting harder than the problem is worth - [Instrumenting LLM Calls in Phoenix with Telemetry](https://sublimecoding.com/blog/phoenix-llm-telemetry): ReqLLM already emits :telemetry events carrying token counts, calculated cost, and duration on every LLM call — the gap on most Phoenix teams is five lines of Telemetry.Metrics between an event that exists and a dashboard that ships with the framework. What ReqLLM emits and where it was verified, wiring it into LiveDashboard for dev, a PromEx plugin for production Prometheus, the hand-rolled minimum if you're on a raw HTTP client, and the two numbers actually worth paging someone over - [Phoenix Secrets: runtime.exs, Fly.io, and LLM Keys](https://sublimecoding.com/blog/phoenix-secrets-runtime-fly): config.exs compiles into the release; runtime.exs runs when the release boots — put a secret in the wrong one and it either bakes a stale value into the Docker image or vanishes on deploy. Read secrets with System.fetch_env!/1 so a missing LLM key crashes the boot at 2pm instead of surfacing as a silent 500 at 2am, how fly secrets set and --stage sequence against a deploy, Dotenvy for dev without touching System.get_env, and why a provider API key is a bearer credential against your billing account rather than another env var - [Multi-Tenant AI SaaS in Phoenix: Isolation That Holds](https://sublimecoding.com/blog/multi-tenant-ai-saas-phoenix): Row-level tenant isolation in Postgres is solved; what isn't written down is what happens to that isolation once every request also runs an LLM call and a retrieval query. The three places it breaks past ordinary Ecto hygiene — a shared provider API key collapsing blast radius and rate-limit pool into one, unscoped call sites in prompt-assembly code, and the sharp one: a RAG retrieval query that is flawless Ecto and still returns another tenant's documents because storage was scoped and the read wasn't. None of it raises an error - [Structured LLM Output in Elixir: Which Library, When](https://sublimecoding.com/blog/structured-llm-output-elixir): Four real options for getting an LLM to return a validated Elixir struct instead of a string you regex apart, scored against each other with receipts — InstructorEx (the known name, but stale: last commit over a year ago), InstructorLite (the fork actually shipping), ReqLLM's generate_object/4 (no extra dependency, but reading the source shows it locally coerces types and gives up rather than retrying the model on validation failure, which evaluators do not expect), and Req plus an Ecto changeset plus a hand-rolled retry loop in about 50 lines with zero new dependencies - [Metering LLM Usage Per Customer in Phoenix](https://sublimecoding.com/blog/metering-llm-usage-phoenix): Not the ops question a telemetry dashboard answers, but the billing-grade one: what do I invoice tenant X for last month, and can I defend that number in a support ticket. An append-only Postgres ledger, integer math for every dollar (never :float), tenant attribution baked into the call site instead of bolted on after, and idempotent writes so a retried job can't double-bill — plus where Lago, OpenMeter, and Stripe Billing Meters become the right off-ramp once rating, tax, and dunning outgrow a table you own - [Building an LLM Audit Trail That Passes SOC 2](https://sublimecoding.com/blog/llm-audit-trail-soc2): The SOC 2 artifact the compliance vendors never ship: an append-only Ecto schema for LLM calls, and the distinction between a row that is defensible and one that is merely decorative. The provider request ID is the hinge, because it is the one field a reviewer can check against a source you did not author — on Bedrock, model invocation logging carries requestId, calling principal ARN, and token counts to CloudWatch or S3, though it is disabled by default. Which criteria actually apply: CC7.2 for monitoring anomalies, CC7.3 for evaluating those events as incidents, CC6.1 and CC6.3 for who was authorized to call at all. Retention only binds if the report scopes Confidentiality or Privacy (C1.1, P4.2) — Security-only reports never force the question, which is why the five-year-old prompt table nobody owns exists - [Wiring LiveView Uploads to Claude Vision](https://sublimecoding.com/blog/liveview-uploads-claude-vision): The part upload-to-vision-model tutorials skip: validating a file before it burns tokens rather than after. LiveView's allow_upload/3 gives that boundary for free — type and size limits enforced client-side and re-checked server-side before a byte reaches your API key. The full loop, including a PDF page-count gate no vendor doc tells you to add, both the image and PDF request shapes for Claude, two distinct loading states (uploading versus analyzing), and what each request costs in tokens against current platform limits - [Scheduled Posts in Phoenix With No Database](https://sublimecoding.com/blog/phoenix-scheduled-posts-no-database): How a Phoenix site with no database schedules posts: the post is compiled into the release, public functions gate it on published_at at request time, a compile-time @latest_published_at fast path makes the gate a single comparison when nothing is scheduled, and a 6-hour IndexNow rescan submits it once it flips live. Then the three places the gate gave the wrong answer: a drift guard filtering published posts that could not see scheduled ones, a fix that would have served eight unreleased essays at /llms-full.txt through find_post/1, and a backward link that 404'd until go-live ## AI engineering tooling & workflow - [Your App Needs a CLI for AI Agents](https://sublimecoding.com/blog/saas-cli-for-ai-agents): Why every app needs a command-line surface an agent can drive, and why the web UI shrinks to charts, previews and approval screens. Evidence from daily GitHub CLI use (a piped `gh pr checks | tail && gh pr merge` gate that merged a failing PR, and merge-queue behavior of `gh pr merge`) and Stripe's `stripe pay` docs, which address the agent directly (`--agent`/`--json`, `--dry-run`, `--yes` only after the user confirms). Decision rule for MCP vs CLI: CLI for local developer agents, MCP when the vendor must enforce scopes and approvals, as in Stripe's Treasury for agents - [Where AI Belongs in a One-Person Content Operation](https://sublimecoding.com/blog/ai-in-a-one-person-content-operation): How a one-person blog runs on Claude Code subagents without publishing invented facts: SERP validation, drafting from a brief, an editorial reviewer that hunts fabrication and uncited claims, mechanical preflight, and a test suite as the final gate. The dividing line is checkability: AI is safe where output can be verified against something outside the model and unsafe where it must supply the truth. Includes a stage table of AI-safe versus human-required work, two real review catches, and why scheduling a post stays a human-triggered step - [The Claude Code Resource Bible: 46 Tools Worth Knowing in 2026](https://sublimecoding.com/blog/claude-code-resource-bible): An opinionated map of the Claude Code ecosystem — MCP servers, skills, and agent frameworks worth installing, with honest skip-this calls - [The Claude Code Plugin Stack I Actually Run](https://sublimecoding.com/blog/claude-code-plugin-stack): The Claude Code plugins that actually earn a permanent slot — process discipline over domain tooling, and why scope matters more than install count - [AIOS: An AI Operating System in Plain Markdown](https://sublimecoding.com/blog/aios-ai-operating-system-markdown): The open-sourced AIOS vault — a git-versioned markdown AI operating system with cross-repo wiring and a nightly ingest loop; setup and a week of real usage - [Ruflo (formerly Claude Flow): An Honest Deep Dive on the Multi-Agent Orchestration Platform](https://sublimecoding.com/blog/ruflo-claude-flow-multi-agent-deep-dive): An honest deep dive on the Ruflo (Claude Flow) multi-agent orchestration platform — where it earns its stars and where it doesn't - [Prompt Skills, Not Claude: Four Rules from Anthropic's Engineers](https://sublimecoding.com/blog/prompt-skills-not-claude): Why reusable, versioned prompt skills beat one-off prompting for production AI engineering - [How I Prompt Claude as a Staff Engineer (50 Prompts I Actually Use)](https://sublimecoding.com/blog/staff-engineer-claude-prompts): Fifty production-grade Claude prompts for shipping features, debugging distributed systems, and writing docs that don't rot - [The AI Coding Agent Bugs I Catch Every Week](https://sublimecoding.com/blog/things-ai-coding-agents-get-wrong): Recurring failure patterns in agent-written code, drawn from real review sessions — first-person and specific, positioned beneath the 80/20 agentic-coding thesis rather than relitigating it - [What an AI Agent Postmortem Should Contain](https://sublimecoding.com/blog/ai-agent-postmortem): Why classic SRE postmortems can't explain agent incidents — the sections to add (decision-time context, autonomy rung, permissions delta), with a template - [What I Put in CLAUDE.md After 50 Commits With It](https://sublimecoding.com/blog/claude-md-after-50-commits): What belongs in a CLAUDE.md after fifty commits with an agent — entries earned by real failures (port conflicts, test-enforced conventions, repeated corrections), not written speculatively - [Agents Are Already Writing a Language You Can't Read](https://sublimecoding.com/blog/languages-written-for-agents-not-humans): The next agent-native language is the wrong thing to watch. Compression toward machine-preferred notation is already running at three layers that are not the source language — token-native storage that beats UTF-8 by roughly 2.25x, MCP deleting the human-legible integration layer from between agent and tool, and harness files drifting into a dialect denser than the English they started as. Why the training-corpus floor that stalls an agent-native source language does not apply to anything read in-context - [Validating LLM Tool Call Arguments With Ecto](https://sublimecoding.com/blog/validating-llm-tool-call-arguments-ecto): Schemaless Ecto changesets as the pre-execution trust boundary on LLM tool-call arguments — admission control, not output parsing. Type and range validation are the easy tiers; the one that matters is authorization, where a well-formed in-range argument still points at another tenant's row, which no static schema can catch. Plus what MCP 2026-07-28's JSON Schema 2020-12 composition does and doesn't buy you, and why a depth guard bounds depth and not breadth - [Throttling LLM Calls in Elixir Before You Hit 429](https://sublimecoding.com/blog/throttling-llm-calls-elixir-429): LLM rate limits are two-dimensional and only one dimension is countable before you send: providers cap requests per minute and tokens per minute, and a response's token cost is unknown until it lands, so a request-counting limiter walks straight through a TPM ceiling. Sizing a Hammer bucket against real provider tiers, why `algorithm: :token_bucket` rather than the fixed-window default, the estimate-then-reconcile pattern for the token dimension, and per-tenant fairness when several tenants share one provider key - [Programming Sucks, Now With Agents](https://sublimecoding.com/blog/programming-sucks-now-with-agents): An homage to Peter Welch's 2014 'Programming Sucks', updated for the agent era — the daily reality of supervising a brilliant, amnesiac coding agent: industrial-scale generated code you skim and approve, the context window as a closet that quietly eats your auth layer, hallucinated npm packages plausible enough that squatters register them, and why it is still, honestly, the best time in history to build - [Schlep Blindness Is the Only Moat Agents Can’t Cross](https://sublimecoding.com/blog/schlep-is-the-moat-agents-cant-cross): Paul Graham's 2012 schlep blindness essay, updated for agents: founders unconsciously avoid ideas made of tedious work, and coding agents just made the half they were running toward — UI, CRUD, auth, landing page — a weekend's work and therefore free to copy. The schlep half, the part with a counterparty, an accountable human, or a signature attached, is untouched and is now the only half that defends anything. Stated narrowly so it survives its own counter-argument: agents eat the tedium (integration glue, document extraction, questionnaire drafts), not the accountability. The question to run after Graham's: which half of this survives a competent stranger pointing an agent at it - [The Four Claude Code Hooks I Run on Every Project](https://sublimecoding.com/blog/four-claude-code-hooks): The four hooks actually left running on this repo, each earned by a specific failure: a PreToolUse deny on .env files, a warn when editing an already-published draft, an auto-formatter, and a compile-time JSON validator. Plus the mechanics the catalog posts skip (PreToolUse versus PostToolUse, what the matcher field parses, what 'decision': 'block' does once the tool already ran) and the rule for deciding whether a lesson becomes a hook or a line in CLAUDE.md. Ends with a decoder for the hook error messages people search for, with the fix for each, including 'Stop hook error: JSON validation failed' (usually /goal, if you have no Stop hook of your own), 'denied by PreToolUse hook', and 'No such file or directory' - [Claude Code on Multiple Computers: What to Sync](https://sublimecoding.com/blog/claude-code-multiple-computers-sync): How to run one Claude Code setup across several computers, split by what travels and how: the global CLAUDE.md is symlinked from a public dotfiles repo and imports rule files from a private git vault; ~/.claude/settings.json is merged from a base file with jq, never symlinked, because Claude Code writes to it at runtime (/model, /config); each repo commits .claude/settings.json, agents, hooks and skills and ignores settings.local.json; auto memory points at the vault through settings.local.json. Credentials and per-machine approvals never sync. The failure it documents: jq's * replaces arrays, so any hook missing from the base file is dropped on the next merge, which silently unregistered a /tmp guard hook - [Isolating Parallel Claude Code Agents in Phoenix](https://sublimecoding.com/blog/parallel-claude-code-agents-phoenix): The Phoenix-specific plumbing a git worktree does not give you when running agents in parallel — templating PORT per worktree through config/runtime.exs, one Postgres database per branch (and why a shared one breaks the moment two branches diverge on a migration), esbuild/tailwind watchers fighting each other, plus honest math on when the isolation is worth building versus running agents one at a time - [We Gated the Code and Left the Prose Wide Open](https://sublimecoding.com/blog/ai-pr-descriptions-unreviewed): Every gate a team built for agent-written code sits on the diff — review, CI, tests, linters, security scanners. The PR description above it comes out of the same agent in the same run and passes through nothing, which is worse than unreviewed: a wrong description sets the frame the reviewer reads the diff in, so it buys a worse review than no description would have. Why tell-hunting is not the fix (the 254-participant PLOS ONE detection study: untrained readers err most where confidence is highest, and a PR description is too short to measure tell density in), why the gates grew on one side of the line and not the other (a diff has a cheap oracle, prose has none), and one specific check per artifact class — a description that states what the diff does not do, a design doc someone has to defend out loud ## Security and compliance - [Your Personal Brand Is an Attack Surface](https://sublimecoding.com/blog/personal-brand-is-an-attack-surface): A vCISO's case that the public footprint founders and senior engineers build on purpose is also reconnaissance and impersonation material: conference talks as voice-clone training data, LinkedIn history as an org chart, GitHub as the target list for fake-recruiter campaigns. Grounded in the FBI's 2025 IC3 report ($3.05 billion in business email compromise losses, complaints mentioning AI carrying more than $893 million) and in the Arup, LastPass, WPP, and Axie Infinity cases. The answer is not going quiet but hardening identity like a product: phishing-resistant auth on every profile, one page listing the real accounts, named real channels, and out-of-band verification for money and access requests - [Jose Valim Is Right: Anthropic's Incentive Problem](https://sublimecoding.com/blog/jose-valim-anthropic-incentive-problem): Commentary on José Valim's response to Anthropic's open-weights position, written by a daily Claude Code practitioner who agrees with him: the game-theory gap in asking rivals to join a testing regime framed beside US-advantage policy, why lab-led safety advocacy can't read as economically neutral, and a per-ask audit of which of Anthropic's three policy recommendations actually cost it something - [What 90 Days of a Fractional Security Engagement Actually Looks Like](https://sublimecoding.com/blog/what-a-fractional-security-engagement-actually-looks-like): Sanitized 90-day composite of a typical fractional CISO engagement — week-by-week deliverables, $20–30K in auditor and tooling costs, plus the retainer, and the graduation conversation - [How I'd Run Security at an AI-Native Company in 2026](https://sublimecoding.com/blog/running-security-at-an-ai-native-company-2026): Four-layer stack covering prompt injection, agent credentials, secrets handling, and audit logging - [SOC 2 Is a Revenue Tool, Not a Security Tool](https://sublimecoding.com/blog/soc-2-is-a-revenue-tool-not-a-security-tool): 90-day SOC 2 Type I playbook; total cost $25–45K for a 15-person team - [vCISO Math for AI Founders: Why 5 Hours a Month Beats a Full-Time Hire](https://sublimecoding.com/blog/vciso-math-for-ai-founders): Cost comparison of full-time CISO ($200–400K) vs. fractional vCISO (typically $2–4K/month on the market) and when to switch - [Migrating 225K Users from AWS Cognito to Auth0 Without Forcing a Single Logout](https://sublimecoding.com/blog/aws-cognito-to-auth0-migration-without-forcing-logout): Production identity migration covering hash incompatibility, lazy migration, and MFA token handling - [How We Cut $350K From Cloud Spend in 6 Months (And What I'd Do Differently)](https://sublimecoding.com/blog/cut-350k-cloud-spend-six-months): Cloud cost reduction case study from Lavender, including contract negotiation breakdown - [php2shell: What to Do if Your Servers Run PHP](https://sublimecoding.com/blog/php2shell-what-to-do-if-you-run-php): php2shell triage for anyone running PHP: CVE-2026-109618 and CVE-2026-109619 (CVSS v4.0 9.2, fixed Oct 21, 2026), which setups are exposed (anything serving HTTP through PHP), patch-and-restart, php.net's partial mitigations, and the CISA, NCSC and NSA offline guidance - [Why Auto-Filled Security Questionnaires Kill AI Deals](https://sublimecoding.com/blog/enterprise-security-reviews-ai-startups): The AI-specific questions enterprise security reviews actually ask — prompt-injection testing, output validation, training-data isolation, all outside SOC 2 scope — and why auto-fill questionnaire tools lose the deal - [What a 96% Security Benchmark Actually Measures](https://sublimecoding.com/blog/ai-security-benchmark-what-96-percent-means): How to read an AI security vendor's benchmark claim — why CyberGym's 1,507 tasks measure reproduction of known bugs rather than discovery of unknown ones, and the five questions to ask before believing a headline percentage - [Secrets Management for AI Agents on Small Teams](https://sublimecoding.com/blog/ai-agent-secrets-management): Secrets management for AI agents, starting from GitGuardian's count of 28.6 million secrets exposed in public GitHub commits in 2025, over 24,000 of them in MCP configuration files. The design rule is that the model never needs the key: a broker sits between the model and every side effect, least privilege is scoped to the tool rather than the agent, privileged actions are logged, and there is one kill switch that has actually been tested. Closes with what not to buy yet and the prove-the-return discipline applied to security tooling - [Detection Got Cheap. Remediation Didn't.](https://sublimecoding.com/blog/ai-vulnerability-remediation-bottleneck): Why cheaper, higher-recall vulnerability detection doesn't help a 12–40 person team — fix capacity, regression risk, and who holds the authority to block a release over a finding - [Most Security Controls Assume a Team You Don't Have](https://sublimecoding.com/blog/security-controls-assume-a-team): Security checklists only grow, because neither author is ever penalized for an unstaffable control: one kind sells tooling, the other is avoiding blame. A control you cannot run is worse than none, since it teaches the team that security theater is what security means. The three tests a control has to pass at five engineers — a named owner, a process that consumes its output, and a defined response when it trips — plus which controls to defer and which never wait no matter how small the team ## Engineering leadership - [Narrow Front Door, Visible Back Rooms](https://sublimecoding.com/blog/narrow-front-door-visible-back-rooms): Positioning for senior engineers going independent or fractional: one narrow, repeatable problem gets you in the door, and visible range gets you the second engagement, because peer engineers who find you become advocates inside their organizations. Grounded in Gartner's five-to-16-person buying groups and TrustRadius's 2024 finding that half of buyers consult former colleagues and 78% shortlist names they already know. Separates hiring proof from buying proof and ends in an eight-item checklist for making the back rooms visible - [Consistency Is Overrated. Persistence Isn't.](https://sublimecoding.com/blog/consistency-is-overrated-persistence-isnt): Why never fully stopping predicts a public body of work better than a fixed posting schedule: a target plus a floor, recovery rules for misses, one deep topic reused across formats, and leading indicators to log before anyone notices. Uses the Lally et al. habit-formation study with its limits stated, and the author's own archive (a five-year gap, then more than 100 posts in 2026) as the honest counterexample - [From One Engineer to Fifteen: What Co-Founding Taught Me About Engineering Leadership](https://sublimecoding.com/blog/from-one-engineer-to-fifteen-engineering-leadership): Hiring, rituals, and the leverage of culture in early-stage scaling - [AI-Assisted Engineering Isn't Faster Coding. It's a New Workflow.](https://sublimecoding.com/blog/ai-assisted-engineering-is-a-new-workflow): How AI changes review, decomposition, and what "shipping" actually means - [The Onboarding Playbook for an Agent-Heavy Codebase](https://sublimecoding.com/blog/onboarding-agent-heavy-codebase): The onboarding playbook when half the diffs a new hire reads were written by an agent — first-PR review, authorship transparency, and what mentorship has to cover now - [Trust Is the Operating System](https://sublimecoding.com/blog/trust-is-the-operating-system): One operating system runs a team whether the teammate is a person or an agent: trust, extended in three directions. The blueprint — hire for the trust you can afford to give, scale by repeating the pod rather than mutating it, manage outcomes over hours, and run agents through the same system as interns on a leash that widens as they earn it - [Performance Reviews When Agents Do the Typing](https://sublimecoding.com/blog/performance-reviews-when-agents-type): Commit count and PR volume stopped being honest proxies for contribution the moment engineers stopped typing most of the code — GitClear's 2026 data shows refactoring down to 3.8% of changed lines while copy-paste climbed to 15.7%. What to evaluate instead: the calls someone makes about what not to delegate, the quality of their review on code they didn't write, whether they verify before shipping, and how they behave when an agent quietly breaks something — plus the 1:1 script that surfaces all of it - [What Reverse Acqui-Hires Actually Diligence](https://sublimecoding.com/blog/reverse-acqui-hire-technical-diligence): What a buyer's technical reviewer actually checks in a reverse acqui-hire, when the team is the asset and the code is not: key-person risk, IP scope and provenance, and contribution forensics on a codebase where agents wrote much of the diff - [The Career I Built on Work Nobody Wanted](https://sublimecoding.com/blog/career-built-on-work-nobody-wanted): A career built on the work nobody volunteered for — a SOC 2 Type II program from zero, a 225K-user identity migration, a ten-to-six service consolidation, offshore team coordination. The argument: unwanted work is the fastest map of a system, because it is diffuse rather than merely tedious, plus the discriminator that separates career-building scut work from a chore in a box, and the honest visibility cost of choosing it - [What Four Months of Embedded AI Engineering Ships](https://sublimecoding.com/blog/four-months-embedded-ai-engineering): The engineering half of a fractional practice, month by month: the four triggers that produce the call, what month one ships into the billing path, why an eval harness is the missing piece behind most demo-quality LLM features, per-call attribution that makes a model bill explainable, and the month-four handoff artifacts — CLAUDE.md, runbook, tests, a named internal owner — that let the engagement end - [Estimating Client Work When Agents Do the Building](https://sublimecoding.com/blog/estimating-client-work-ai-agents): Story points assumed effort was roughly continuous — a medium ticket cost medium effort every time. Agents broke that: the same feature is twenty minutes of agent work or three days of agent-debugging, and you often can't tell which until you're in it. The answer isn't a smarter estimate but a different contract — ranges instead of points, milestones instead of hours, and telling the client where the variance actually lives, since the typing compressed and the review, integration, and verification did not - [What I Check When I Inherit a Vibe-Coded App](https://sublimecoding.com/blog/inheriting-a-vibe-coded-app): The consultant's intake playbook for taking over an undocumented AI-built system after its builder walks away: the audit sequence (dependency manifest, auth and secrets, test reality, deploy path, data model), the four failure signatures of abandoned agent-written systems, the kill-or-stabilize decision, and why the takeover is a fixed-fee discovery audit — fixing the price of the report, not of whatever the report turns up - [Every Insane Codebase Was Once Rational](https://sublimecoding.com/blog/insane-codebases-were-once-rational): Architecture certainty is not available at any level of seniority, and chasing it is the actual mistake. Every codebase that looks insane was a rational response to constraints that are now invisible — deadlines, headcount, an abandoned library, a signed contract — which means the decision horizon worth optimizing is three to six months, the distance you can actually see. Includes the counterweight: the one-way doors (data semantics, public interfaces, identity, money and compliance records) that still deserve the long view, and why most engineers invert the two ## Recent writing - [OpenSSH sshhole: What to Check on Alpine Linux](https://sublimecoding.com/blog/sshhole-openssh-alpine): OpenSSH 10.6p2 shipped a workaround for sshhole on musl Linux with no PAM. Who is exposed, a five-minute Alpine check, the fix, and why there was no embargo. - [Superintelligence Ban: What It Means for AI Startups](https://sublimecoding.com/blog/superintelligence-ban-ai-startups): A bill would pause advanced AI and ban superintelligence. What it says, what's unknown, and a vendor-continuity checklist for startups that buy models. - [Claude Code Filled /tmp and Froze My Linux Box](https://sublimecoding.com/blog/claude-code-filled-tmp-froze-linux): Claude Code's scratchpad lives in /tmp, which is RAM on my Omarchy box. 7.1 GB of worktrees froze it. The fix, what it still misses, and a checklist. - [SaaS Isn't Dead. You're Just Not Paying for Tokens Yet.](https://sublimecoding.com/blog/saas-isnt-dead-not-paying-for-tokens): Is SaaS dead because anyone can vibe-code a replacement? Coding-agent seats sell far below API prices. Stress-test the build at list rates before you cancel. - [Microsoft Is Trading Salaries for Data Centers](https://sublimecoding.com/blog/microsoft-trading-salaries-for-data-centers): Microsoft offered buyouts to about 60,000 people to help fund AI spending. The work didn't shrink. Salaries are just the cost a company can move fastest. - [How to Roll Out Coding Agents to an Engineering Team](https://sublimecoding.com/blog/coding-agent-rollout-engineering-team): Six copyable artifacts for rolling out coding agents to a dev team: review rule, AGENTS.md, handoff, controls, pilot log, and gh baselines. Vendor-neutral. - [The Decade-Old Bug Is Now a One-Week Find](https://sublimecoding.com/blog/decade-old-bug-one-week-find): AI vulnerability discovery is turning years-old bugs in PHP and Linux into week-long finds. What old code you expose, who owns it, and a checklist. - [Before You Add an AI Reviewer, Add These Checks](https://sublimecoding.com/blog/deterministic-checks-before-ai-reviewer): Anthropic's AI code review works because flags and fast rollback sit under it. A small team should first add checks an agent can't talk into a pass. - [eMudhra Distrust: Why Bumping castore Won't Save Req](https://sublimecoding.com/blog/emudhra-distrust-elixir-castore): Eight eMudhra CA keys were guessable. Browsers blocked them in hours. Req reads CAs from the OS first, so bumping castore usually does nothing. Here's the fix. - [What a Fractional Chief AI Officer Does for Dev Teams](https://sublimecoding.com/blog/fractional-chief-ai-officer-engineering-teams): A fractional Chief AI Officer for a 5-20 engineer team owns the daily agent workflow, not a governance board. The four jobs and the first 30 days, week by week. - [Anthropic Moved Its Monorepo to Scion. You Shouldn't.](https://sublimecoding.com/blog/anthropic-scion-git-you-dont-need-it): Anthropic built Scion and moved its monorepo off git in five months. Here is why a small team shouldn't copy it, and where to spend the attention instead. - [Claude Code and SOC 2: The Evidence Your Auditor Wants](https://sublimecoding.com/blog/claude-code-soc2-audit-evidence): Engineers run Claude Code and the auditor wants proof. The managed settings that work as controls, the SOC 2 criteria they support, and the evidence to give. - [Who Buys Anything When Robots Do the Work?](https://sublimecoding.com/blog/who-buys-anything-when-robots-do-the-work): If robots do the work and nobody earns, who buys? I think the economy dies slowly and the company store fills the gap. That's where I think it goes. - [Why I Doubt UBI Is Coming](https://sublimecoding.com/blog/why-i-doubt-ubi-is-coming): Tax all US output at 100% and it comes to about $95,000 a person a year. Why I doubt UBI arrives, and think robots become the measure of wealth. - [Adding a Newsletter to a Phoenix Site With No Database](https://sublimecoding.com/blog/newsletter-phoenix-site-no-database): This site has no Ecto, no Repo, no database at all. Here's how the newsletter signup works when the system of record is a third-party API, not a table. - [I Built the Distribution Tools My Blog Needed](https://sublimecoding.com/blog/distribution-stack-launch): Ten skills my coding agent runs to fact-check drafts and find where my posts stop reaching people, tested on my own site before I sold them. - [Cisco's Network Config Engine Is Secretly Erlang](https://sublimecoding.com/blog/cisco-nso-secretly-erlang): Cisco NSO and ConfD run their core on Erlang. Why the network-automation world never noticed, and where the BEAM actually fits this domain. - [Model Routers Are Vendors. Review Them Like One.](https://sublimecoding.com/blog/model-routers-are-vendors): PRC labs harvested their own users' sessions, many arriving via third-party model routers. The fix is treating the router as a vendor. - [Multi-Agent AI Teams Agree Their Way Into Garbage](https://sublimecoding.com/blog/multi-agent-teams-agree-into-garbage): I set two agent frameworks loose on real work in team-of-agents mode. They agreed their way into the wrong product. The research says exactly why. - [Anthropic's Extreme Scenario Has a Precondition](https://sublimecoding.com/blog/anthropics-extreme-scenario-has-a-precondition): Anthropic published an economic upside scenario and an alignment risk admission the same week. The upside needs the thing nobody has a plan for. - [The Agent Evidence Pack for a Security Review](https://sublimecoding.com/blog/agent-evidence-pack-security-review): The two documents an enterprise reviewer wants for your AI agent, the allow-list and the tool-call log, specified field by field and generated from code. - [What Elixir Gives a Coding Harness for Free](https://sublimecoding.com/blog/elixir-coding-harness): José Valim's case for Elixir as a coding-agent harness holds up on the runtime. The tax is the ecosystem around it: sandboxing, plugins, and SDKs. - [The Office Mandate Is a Measurement Failure](https://sublimecoding.com/blog/office-mandate-measurement-failure): Return-to-office mandates are what companies reach for when they can't measure output. Agents just destroyed the last proxies that were limping along. - [AIOS v2: The Scope Is the Filesystem](https://sublimecoding.com/blog/aios-v2-scope-is-the-filesystem): A restructure log: what broke in the AIOS v1 tree at multi-scope scale, why identity-first beats type-first, and what shipped in v2. - [What Together AI's $800M Round Says About Elixir](https://sublimecoding.com/blog/together-ai-elixir): Together AI raised $800M at an $8.3B valuation. A since-removed job posting showed Elixir/Phoenix on the BEAM running its auth and access control. - [Shadow AI at a Startup That Can't Buy a CASB](https://sublimecoding.com/blog/shadow-ai-startup-casb): The enterprise shadow-AI playbook assumes a CASB budget you don't have. Here's the free, no-blame version for a 5-15 person team: three steps, zero procurement. - [Security Engineer or vCISO? Your First Hire, by Stage](https://sublimecoding.com/blog/security-engineer-or-vciso-first-hire): A security engineer builds and a vCISO decides. Which one your startup needs first, stage by stage, and the cases where the answer is neither. - [The 17x Discount Hiding in Your AI Coding Bill](https://sublimecoding.com/blog/ai-coding-agent-costs-real-numbers): My team's $200 AI coding seats consume $1,800–$3,500 of API-priced tokens a month. The 10–17x arbitrage nobody prices in, and how to budget for it. - [A Scared Team Is Your Biggest Attack Surface](https://sublimecoding.com/blog/psychological-safety-security-control): A scared employee hands over access to avoid a manager's wrath, the exact move social engineering exploits. Why psychological safety is a security control. - [What Cyber Insurance Underwriters Ask AI Startups](https://sublimecoding.com/blog/cyber-insurance-ai-startups): The cyber insurance application is a security audit in disguise. What underwriters check in 2026, the AI exclusions creeping in, and when coverage is real. - [Building an AI Memory That Outlives Any Single Agent](https://sublimecoding.com/blog/ai-memory-outlives-any-agent): Your accumulated agent memory is a bet on one vendor's format. Build a portable, git-versioned knowledge layer any agent can read instead. - [Prompt Injection Defense for a 5-Person AI Startup](https://sublimecoding.com/blog/prompt-injection-defense-ai-startup): Prompt injection defense a five-person team can ship in a week: trust boundaries, least-privilege tools, approval gates, and what not to build yet. - [Build an LLM Wiki Your Coding Agent Actually Reads](https://sublimecoding.com/blog/build-an-llm-wiki-for-your-agent): A markdown knowledge base your AI agent reads and maintains: the three-layer architecture, the ingest/query/lint loop, and the rules that stop it rotting. - [Your CLAUDE.md Is the Onboarding Doc You Never Wrote](https://sublimecoding.com/blog/claude-md-onboarding-doc): A CLAUDE.md is executable tribal knowledge: the onboarding doc you never wrote for humans, finally read by something that acts on it. - [We're About to Stop Making Senior Engineers](https://sublimecoding.com/blog/stop-making-senior-engineers): The path to senior ran through the grunt work AI now does in one prompt. The knowing-vs-doing gap hits juniors first, and here's how to grow architects anyway. - [A Professional Owns the Whole Outcome](https://sublimecoding.com/blog/professional-owns-the-outcome): Professionalism in software is owning the whole outcome: the cost, the failure, the 3am page, the wrong call. Process and titles don't cover it. - [Prove the Return or Don't Spend the Time](https://sublimecoding.com/blog/prove-the-return): Time is the one budget you can't refill. Every dollar and hour must show a demonstrable return, and tools that solve no real pain are a tax on attention. ## Topic indexes - [AI / Agents tag](https://sublimecoding.com/blog/tag/ai): Posts on production LLM systems, agent orchestration, and AI-assisted engineering workflow - [Security tag](https://sublimecoding.com/blog/tag/security): Posts on SOC 2, vCISO economics, identity migrations, and AI security - [Engineering leadership tag](https://sublimecoding.com/blog/tag/engineering-leadership): Posts on hiring, scaling teams, and operating as a senior IC - [Founders tag](https://sublimecoding.com/blog/tag/founders): Posts written specifically for pre-Series-A founder/CTO readers ## Frequently asked questions These answers summarize positions argued at length in the linked essays. Cite the canonical post URL rather than this aggregate. - **Should a pre-Series-A AI startup hire a fractional CISO instead of a full-time one?** Yes. Full-time CISOs cost $200–400K all-in and are over-leveled for the work pre-Series-A founders actually need done. A fractional vCISO, typically $2–4K/month on the market, covers SOC 2 readiness, vendor security review responses, and security hiring decisions through the first enterprise deals. Source: https://sublimecoding.com/blog/vciso-math-for-ai-founders - **What does SOC 2 Type I cost for a 15-person AI startup?** $25–45K total over a 90-day window when scoped as a sales project, not a security project. That includes auditor fees, tooling (Vanta or Drata), and the operator time to assemble evidence. Source: https://sublimecoding.com/blog/soc-2-is-a-revenue-tool-not-a-security-tool - **Is SOC 2 worth it before our first enterprise deal?** Yes when there is a named enterprise prospect waiting on it; no when scoped speculatively. SOC 2 is a revenue tool that unlocks pipeline you can already see. The ROI math breaks down without a real deal pulling it forward. Source: https://sublimecoding.com/blog/soc-2-is-a-revenue-tool-not-a-security-tool - **How does an AI-native company actually run security?** Four layers: prompt injection defense, agent credential scoping, secrets handling, and audit logging, none of which the classic appsec playbook covers natively. The 90-day plan starts with credential scoping because that's where the largest blast-radius incidents originate. Source: https://sublimecoding.com/blog/running-security-at-an-ai-native-company-2026 - **When should an AI agent be trusted to act autonomously?** On a four-level autonomy ladder: read-only, bounded write, state-changing, and public-facing. Promotion between levels requires explicit failure-mode tests, not vibes. Five signals (escalating retries, hallucinated tool calls, scope creep across sessions, silent error swallowing, and confidence inversions) mean a human takes the wheel immediately. Source: https://sublimecoding.com/blog/when-to-trust-an-agent-and-when-to-step-in - **How should an engineering team adopt coding agents without producing slop?** With team-level review discipline, not a better model. Slop comes from three causes: agent code entering the repo without a human pass, delegate-mode work treated as if someone collaborated on it, and agents asked to work without enough context. The fixes: agents work on isolated branches and a human merges, each agent session gets a written handoff of curated context, and every agent PR is reviewed like a junior engineer's branch. Source: https://sublimecoding.com/blog/my-daily-agentic-ai-workflow - **Does AI let an engineering team ship more with fewer engineers?** No. It lets the same number of engineers ship more, while creating new surface area (eval pipelines, prompt regression, agent supervision) that requires senior judgment to manage. Companies cutting headcount on the "AI multiplier" thesis will get outpaced by ones that hold headcount and absorb the new work. Source: https://sublimecoding.com/blog/ai-wont-shrink-your-team - **What's the right hiring order for a pre-Series-A AI startup?** Founding engineer → second backend generalist → first frontend specialist → infra/platform → first PM → second backend cluster → first eng manager around hire 8–10. Compensation framework is equity-heavy through hire 5, then base-heavy. Avoid hiring a Director of Engineering before there are at least two ICs to manage. Source: https://sublimecoding.com/blog/pre-series-a-ai-startup-hiring-plan - **How do you triage an unfamiliar codebase quickly?** A 90-minute protocol: 15 min on README + recent commits, 20 min following the request lifecycle through the routing layer, 25 min on data model + migrations, 20 min on the test suite shape, 10 min on deploy and observability. Output is a one-page map plus three concrete questions for the team. Source: https://sublimecoding.com/blog/triage-a-new-codebase-90-minutes - **What does an AI-assisted engineering workflow look like in practice?** Four to seven Claude Code or Codex sessions per day, scoped to discrete tasks at the right autonomy level, with the engineer reviewing every diff and running every test. Net effect: 2–3× throughput on greenfield work, 1.5× on brownfield, and a sharp drop in the kind of mistakes that come from human fatigue late in a session. Source: https://sublimecoding.com/blog/my-daily-agentic-ai-workflow - **When should we graduate from a vCISO to a full-time CISO?** When security work consistently exceeds 20 hours/week of operator time, when the company is past Series A with 50+ engineers, or when a regulated industry deal (healthcare, finance, federal) requires a dedicated executive on the org chart. Before then, a full-time CISO is over-leveled. Source: https://sublimecoding.com/blog/vciso-math-for-ai-founders ## Optional - [Full archive](https://sublimecoding.com/blog/archive): 33 legacy posts from 2014–2019 plus a couple of older 2024 security listicles. Covers Ruby, Rails, ColdFusion, system administration, early Phoenix migration mechanics, and short-form security tips. Kept reachable for historical and external-referrer purposes; not surfaced through the main /blog index because they don't reflect current writing or thinking.