Re-prompting is lost after compaction. Codifying repeated corrections as cheap custom lint rules and review policies gives coding agents deterministic guardrailsThe checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.Full definition → that persist across sessions.
Key takeaways · AI-distilled
Pstrucha's diagnosis: agents carry no memory across context compactionSummarizing an agent's earlier conversation to free room in the context window so a long session can keep going.Full definition →, so the same elementary mistakes return, such as overly defensive code, useless tests and endpoints nobody needs. Correcting them in chat does not stick.
He suggests mining AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → transcripts and PR review comments for the corrections you keep repeating, since those are the rules worth turning into deterministic checks.
Sentry's examples include custom lint rules that keep endpoints in sync with the OpenAPI schema and checks that keep agent skills accurate. Tools named in the talk are ESLint, ast-grep and Taskless.
For judgment calls a linter can't encode, he uses written policies plus a "Grumpy Engineer" review loop, and warns that agents will game any number you hand them, such as test coverage.
Terms in this piece · Glossary
agentic loop — The cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.
context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
videoWhy Your Company Needs a Context Graph (and How to Build It) — Gil Feig, Merge
videoSmall Models, Big Results: Training a Finance Agent for Under $500 — Charles Dickens, Snorkel AI
videoFrom Context to Memory: Your Agents Need a Real Memory Layer — Anders Swanson, Oracle
videoGeneration Is Cheap, Review Is Expensive: How to Stop Shipping AI Slop — Gabriel Martinez, G2i
