What an AI Agent Postmortem Should Contain
Classic SRE postmortems can't explain agent incidents. The sections to add (decision-time context, autonomy rung, permissions delta), with a template.
We can't find the internet
Attempting to reconnect
Something went wrong!
Attempting to reconnect
Writing · Tag
56 posts on agents, page 4 of 5. Or browse the full writing index →
Classic SRE postmortems can't explain agent incidents. The sections to add (decision-time context, autonomy rung, permissions delta), with a template.
Your accumulated agent memory is a bet on one vendor's format. Build a portable, git-versioned knowledge layer any agent can read instead.
I shipped more code in 2026 than the previous four years combined. The commits are real. The productivity is real. What I lost is harder to measure.
An homage to Peter Welch's Programming Sucks, updated for agents: a genius intern with amnesia, hallucinated packages, and a closet that eats your auth layer.
Most AI agent frameworks reinvent a job queue badly. Oban already is one: durable, idempotent, retry-aware. Run an agent loop that survives a deploy mid-run.
Amazon mandated AI coding, let it touch infrastructure unwatched, and lost millions of orders. The fix was more humans per deploy.
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.
A CLAUDE.md is executable tribal knowledge: the onboarding doc you never wrote for humans, finally read by something that acts on it.
Eight failure patterns I see running AI coding agents daily: the confident wrong answers, the lost context, and the bugs they reliably ship.
Every AI agent framework runs the same loop: observe, decide, act, repeat. Here it is in 50 lines of Elixir with no framework, just a GenServer.
Go won the agent daemon and Python owns the reasoning. The layer nobody claimed is the one that bites you: thousands of long-lived, stateful, crash-prone agents you must keep alive. That's the BEAM's home turf.
Most engineers prompt Claude one sentence at a time. Anthropic's own engineers prompt skills. Four rules from their recent talks, with the operator nuance the talks left out.