Agent Operating Notes

Short operating notes from redacted Claude and Codex sessions: boundaries, prompts, failure modes, and team practices for AI-assisted engineering.

Restart AI Coding Sessions When Context Stops Being Verifiable

Restart an AI coding session when context stops being verifiable. Use a checkpoint that preserves facts, uncertainty, and the next safe probe.

Before an AI Agent Asks a Question, Make It Search

Use AI agent context gathering to reduce avoidable team interruptions. Separate cheap lookups from decisions that genuinely need a human.

Make Agent Integration Failures Explicit Before Asking Humans

Classify AI agent integration errors before asking humans for context. Preserve missing input, unavailable capability, denied access, and stale sources.

Define AI Agent Work as Verifiable Execution Slices

Use AI agent task decomposition built around one outcome, changed surface, proof path, rollback, and stop condition instead of arbitrary small tasks.

Turn Painful Agent Runs Into Evaluated Skills

Build AI agent skills from hard-won runs by extracting invariants, bounding authority, and testing new cases instead of saving an unverified prompt.

Give AI Agents Permissions by Intent, Not Identity

Scope AI agent permissions to inspect, plan, patch, deploy, or verify. Review authority transitions instead of granting broad access by assistant identity.

Build a Provenance Ledger for AI Agent Research

Use an AI agent provenance ledger to separate observed, inferred, contested, and unsupported claims before research drives engineering decisions.

Use Reversible Rollouts for AI Agents in Production

Run an AI agent production rollout through observe, prepare, shadow, limited, and general stages with evidence, thresholds, and executable rollback.

Monitor Long-Running AI Agent Work Without State Drift

Design AI agent monitoring as a read-only delta heartbeat with stable evidence, idempotent checkpoints, anomaly thresholds, and terminal conditions.

What a Genuinely Blocked AI Agent Should Report

Use an AI agent blocker report to stop equivalent retries, preserve completed evidence, identify the failed dependency, and name the smallest unblock.

Teaching an AI Reviewer to Disprove Itself Before It Comments

AI review false positives fall when do-not-flag policy, falsification checks, and trace reporting ship as one system.

Typed Intent Routing Before an Agent Touches the Broker

Typed intent routing makes model-assisted agent routing usable only when scope, schema, approval, and audit evidence gate execution.

Choosing Real-Time Commit Story Mining Over Bulk GitHub Indexing

Commit story mining works better when agents observe intent in real time instead of indexing every repository change first.

When Editorial Quality Scores Stopped Being the Right Sort Order

AI editorial workflow boards should separate freshness, quality, and actionability so reviewers see new candidates first.

Mining Huge AI Session Transcripts Without Drowning in Them

AI transcript mining needs evidence-first search, narrow excerpts, and claim labels before large agent logs become source material.

Gate default model changes with golden sessions

Golden sessions give AI runtime model default changes a regression contract across prompts, profiles, approvals, CI, and live evidence.

When Verification Schemas Drift From Runtime Reality

Verification schema drift happens when agent output schemas and runtime validators diverge. Make the contract executable before live runs.

Turning a Pilot AI Platform Into an Operable Service Contract

An AI service contract turns a working internal AI pilot into an operable service with owners, dependencies, and fail-closed gates.