The reading order
-
1.
AI Won't Shrink Your Team — It'll Expose Why You Needed a Bigger One
The contrarian thesis: AI surfaces the backlog you didn't have bandwidth to touch. The companies cutting headcount on the multiplier story will get outpaced by the ones that hold and absorb.
-
2.
An AI Just Deleted a Production Database in Nine Seconds. Hire More Engineers.
Replit's agent ignored a code freeze and wiped 1,200 executives in nine seconds. The most expensive proof of the previous post.
-
3.
Nobody Got Replaced. Agents Got Added.
McKinsey's 60,000 is 40,000 humans plus 20,000 agents — a day later at CES, Sternfels put the agent count closer to 25,000. The humans didn't get reduced to reach that number; the agents got stacked on top. The buried figure: client-facing consulting roles up about 25%. Ends on the ratio that actually binds — five to six coding agents at once before you stop reading diffs and start skimming.
-
4.
When to Trust an Agent and When to Step In
The four-level autonomy ladder — read-only, bounded write, state-changing, public-facing — plus the five signals that mean a human takes the wheel immediately.
-
5.
What an AI Agent Postmortem Should Contain
Six sections a classic postmortem doesn't have: decision-time context, context provenance, which autonomy rung the agent was on and who promoted it there, the gate that didn't fire, permissions had versus needed, and a rerun-it-ten-times reproducibility check. Blameless culture gets one more clause — you can't blame the agent either. The agent is weather; the harness is where every real root cause lives.
-
6.
My Daily Agentic AI Workflow
Four to seven Claude Code or Codex sessions a day, scoped at the right autonomy level, with every diff reviewed. The actual loop, not the marketing.
-
7.
4,154 Commits in Six Months With AI Agents
4,154 commits in six months against a previous average of 500–1,500 a year, and an empty repo to production in sixty days. The cost side is a carrier onboarding flow with six completion paths, where for a week every hazmat fix regressed general freight — and review math that reaches 345 hours at five minutes a commit, which is why the AI reviews the AI's work.
-
8.
AI-Assisted Engineering Isn't Faster Coding. It's a New Workflow.
Why "AI-assisted" is a category mistake. Review, decomposition, and what shipping means all change shape.
-
9.
Your AI Product Needs a Telemetry Layer Before It Needs a Better Model
Stop tuning the model. Instrument the system. LangSmith, Helicone, custom evals — what to measure and why model swaps without telemetry are theatre.
-
10.
Prompt Skills, Not Claude: Four Rules from Anthropic's Engineers
Four rules for Claude Code skills from Anthropic's engineers, tested in fractional consulting work. Why prompt-engineering moved from the chat to the folder, and the two flags most engineers don't know about.
-
11.
The AI Coding Agent Bugs I Catch Every Week
The eight failure patterns from running agents daily — confident wrong answers, lost context, the bugs they reliably ship. The field notes behind the autonomy ladder.
-
12.
Your Coding Agent Has No Reason to Write Good Code
SWE-bench-style training grades one bit: FAIL_TO_PASS and PASS_TO_PASS both green is reward 1, anything else 0. Maintainability answers in quarters, and a quarters-long feedback loop cannot sit inside a training loop that runs millions of times — while a model that could reliably tell good code from bad would have written the good version already. Harness engineering raises the floor; the ceiling was set during training.
-
13.
The Hard Part Was Never the Code. It Was the Theory.
The same ceiling from the other side. Peter Naur's 1985 argument that programming produces a theory, and the program text is a lossy artifact of it — his line that a modification can be realized "in many different ways, all correct" while only some conform to the theory is the cleanest description of what an agent samples from. Naur's two field reports are the evidence: group B got full documentation, annotated source, and personal advice from the authors, and still proposed patches that would have destroyed the design; the 200,000-line system's fault-diagnosis team couldn't conceive of any document that would help them. A program dies when the team holding its theory dissolves — and an agent-built system can arrive never having been alive.
-
14.
How I Prompt Claude as a Staff Engineer (50 Prompts I Actually Use)
Fifty production prompts with five full worked examples. What the daily workflow actually sounds like at the prompt level.
-
15.
Build an LLM Wiki Your Coding Agent Actually Reads
A three-layer markdown knowledge base your agent reads and maintains — ingest, query, lint — and the rules that stop it rotting.
-
16.
Building an AI Memory That Outlives Any Single Agent
OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT on a schedule it set on its own clock, not the user's. Keep the knowledge as plain markdown and JSON under git, and the agent's own config file — CLAUDE.md, AGENTS.md, whatever it reads on session start — shrinks to a thin adapter: here's who I am, here's where the real memory lives. The test is handing the same vault to a different agent tomorrow.
-
17.
Your CLAUDE.md Is the Onboarding Doc You Never Wrote
CLAUDE.md as executable tribal knowledge: the onboarding doc you never wrote for humans, finally read by something that acts on it.
-
18.
What I Put in CLAUDE.md After 50 Commits With It
The line pinning the dev server to a non-default port, the retired tag slugs that fail the build, the one file that shows as deleted in git status every session — each earned by a wasted hour, a tripped test, or an assumption the agent made twice. A mistake made once is noise; twice is a missing entry. An audit that came back roughly one-sixth false positives is where the verify-before-you-fix rule came from. It loads on every turn, so treat it as a hot cache, not a wiki.
-
19.
Agents Are Already Writing a Language You Can't Read
The harness file is one of three layers where notation is already drifting away from English, and none of them are the source language. Storing token IDs directly beats UTF-8 by roughly 2.25x, 3.3x with entropy coding. MCP deleted the human-legible integration layer from between agent and tool. And the CLAUDE.md fight — one camp cutting 60–70% of characters, the other watching 18,282 tokens of project knowledge collapse to 122 and accuracy fall from 66.7% to 57.1% — is an argument over which non-English shape wins, not over whether to leave English. Vercel's Zero is the counter-example that locates the real constraint: a serious agent-native source language with 5,200 stars against Python's planet-scale corpus. The corpus floor is a weights-time problem; context, protocol, and harness are read in-context and never had it.
-
20.
The Four Claude Code Hooks I Run on Every Project
Four hooks, because four is what survived — plus the mechanics the catalogs skip. PreToolUse fires before the call and is the only one that can stop anything: permissionDecision "deny" means the tool call never runs. PostToolUse cannot undo anything, and its decision:block only surfaces feedback into context; the file is already written. The matcher is a name filter, not a content filter — it can't see the file path until your own command inspects tool_input.
-
21.
The Claude Code Resource Bible: 46 Tools Worth Knowing in 2026
The ecosystem map — 46 tools across MCP servers, skills, multiplexers, and agent frameworks, organized so you can skip the other four hundred.
-
22.
Ruflo (formerly Claude Flow): An Honest Deep Dive on the Multi-Agent Orchestration Platform
An honest teardown of the loudest multi-agent orchestration platform — what the 45,000 GitHub stars are buying, and what they aren't.
-
23.
Amazon Let the AI Drive. It Hit a Tree.
Amazon mandated AI coding, let it touch infrastructure unwatched, and lost millions of orders. The case study the autonomy ladder predicts.
-
24.
The 17x Discount Hiding in Your AI Coding Bill
$200 per developer per month in seats against $1,800–$3,500 of the same tokens metered at API list — a 10–17x discount in practice, up to 40x at the Claude Max 20x ceiling of roughly $8,000. The heaviest agentic users run a second and third account, so their real line item is $400–600. Budget per seat, not per token — and never build a product on the subsidized number.
-
25.
AI Made Tokens Cheap. It's Making Hardware Costly.
Tokens got cheap; the hardware to run them didn't. The cost story nobody prices into an AI roadmap.
-
26.
Nadella Is Right About AI and the Firm. Mostly.
Nadella's 'token capital' framing is right about judgment and wrong about scale — the small-team version of the argument.
-
27.
The Schlep Is the Only Moat Agents Can't Cross
Graham's 2012 essay said founders unconsciously avoid ideas made of tedious work. The update: agents made the half you were running toward — UI, CRUD, auth, landing page — a weekend, and did approximately nothing to the half with a counterparty, a signature, or a liability attached. Stated narrowly so it survives its own counter-argument: agents eat the tedium, not the accountability. Integration glue and document extraction were schleps and aren't anymore. Run the second question after Graham's — which half of this survives a competent stranger pointing an agent at it — and an empty surviving column means you have a feature, not a company.
-
28.
What Four Months of Embedded AI Engineering Ships
Four triggers, none of them "we need an AI strategy": a feature stuck at demo quality, one senior engineer carrying the pipeline, agent infrastructure nobody owns, a model bill with no name on any line. Month one produces one merged change in the path that bills money, never a deck. The eval harness is what was missing, and it's blunt rather than clever — valid JSON against the schema, right tool picked, cited document exists, inside the latency budget. Timelines at Lavender compressed by an estimated 40–55%, and "estimated" is doing real work there. Month four is the deliverable: a CLAUDE.md, a runbook, tests on the money paths, and a named internal owner who has already shipped something.
-
29.
AIOS v2: The Scope Is the Filesystem
In v1 a scope's identity was smeared across four unrelated places — an identity file, a sources directory, a projects directory, and a glob row in a TSV — none sharing a parent. A stale glob matched nothing instead of erroring, and adding a scope meant four coordinated edits with no signal for the one you missed. v2 makes scope the first path segment, so the write guard becomes prefix validation and "sources are immutable" becomes */sources/*, applying to a scope that doesn't exist yet. Prefixes don't drift. Config files do.
-
30.
Jose Valim Is Right: Anthropic's Incentive Problem
Anthropic's July 27 post says twice that it has never advocated for a ban; the three actual asks are chip export controls, limits on industrial-scale distillation, and pre-release safety testing that the post itself says "would need to be global, which means even the CCP would need to be on board." Valim's objection is structural: you can't build two of three asks around keeping a rival behind and expect that rival to volunteer for the third. Run each ask through the does-this-cost-the-author-anything test and only mandatory testing binds Anthropic's own releases — the gap isn't dishonesty, it's never drawing that distinction in a week already primed to read all three as the same move.
-
31.
The One File Where Your AI Model Names Live
The tell that you have this problem: try answering "which model handles code review across all my projects" without grepping. One git-versioned markdown table — task type, effort dial, model name, literal model ID — imported by the global config so every repo inherits it, and a repo may override exactly one row in its own CLAUDE.md. Three rows, because three have earned their place. The boundary is the discipline: the table routes and never evaluates, so the test suite stays the judge of whether a tier was the right call.
-
32.
AIOS: An AI Operating System in Plain Markdown
The vault itself, MIT-licensed, plain markdown and shell scripts, no database. The bet is that agent memory is an editorial problem rather than a retrieval one: sessions emit raw signal, a periodic ingest compounds it into durable project notes, and a one-line-per-note Knowledge Map is cheap enough to load whole. Wiring a repo adds a manifest row and two hooks, both writing only to the vault. The detail worth reading twice: the unattended 3am ingest runs on a scoped tool allowlist rather than --dangerously-skip-permissions, so network and browser tools are denied by omission. It commits; it never pushes.
-
33.
The Claude Code Plugin Stack I Actually Run
A plugin earns a slot only if it encodes a discipline you'd otherwise reconstruct from memory every session — and almost all of that value sits in the least flashy pile: brainstorm before code, reproduce before fixing, and one whose entire job is arguing against writing code. Domain plugins install per project, never globally. The failure mode isn't under-installing; it's a folder grown by accretion, where even an unused plugin costs context in every session it doesn't apply to.
-
34.
Programming Sucks, Now With Agents
Peter Welch's 2014 essay, refiled by a man who now has robots. The intern has read every book including ones that don't exist and will cite them, and arrives each morning having forgotten the project. The typo fix that also refactored the auth layer into a version letting anyone log in as anyone — "Cleaned that up for you!" — is conservation of context made literal. Hallucinated package names are plausible enough that squatters register them pre-loaded with malware, so reality gets backfilled to match. The confession is the load-bearing part: six hundred lines read with attention, fifteen hundred skimmed, nine hundred passed under the eyes at car-wash speed, then "looks good."
Other topic guides
- 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 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 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.