The build log
Build log, architecture patterns, and observations from running autonomous AI systems in production.
Defer a reversible capability with a named, measurable trigger. Pull a decision forward when waiting expands security, privacy, compliance, data-integrity, contractual, or migration risk. Good timing is an explicit policy, not a hunch.
As parallel work increases, repeatedly reconstructing context becomes a real cost. Durable notes and handoffs can help when they are scoped, classified, current, access-controlled, and used as evidence for human judgment rather than as automatic instruction.
Why the pipeline component that writes your published output needs the tightest lockdown: no shell, no network, draft-only writes, and typed fields for untrusted text.
A single debug flag set to false in October 2023 rode a parked branch for two and a half years, merged in April 2026, and silently dry-ran every admin write in production. The postmortem is about how scaffolding debt outlives memory and what we changed to prevent it.
When a pipeline reuses its own artifacts as sources, publication-time redaction alone cannot govern sensitive data. Privacy requires classification, minimization, protected raw sources, de-identified derivatives, output review, retention controls, and monitored exceptions.
Claude Code logs every session to undocumented JSONL. Three parser traps: launch-cwd directory names, tool results labeled as user turns, and the nearly absent aiTitle.
Automated reviewers can identify real risks while a sequence of locally reasonable edits harms the whole artifact. The answer is not to waive safety findings, but to use calibrated review, minimal changes, whole-artifact rereads, and-where appropriate-direct human input or non-publication.
A broken paywall flow left a credit count reachable only by raw UPDATE. The safer move is a twenty-minute staff tool that reads, dry-runs, validates, caps, and logs.
A release workflow can use a remote control plane rather than a local checkout, but only when protected branches, approval evidence, provenance, deployment checks, and post-deployment verification remain explicit.
A refactor silently dropped a class of records while the rendered file still looked plausible. A large diff against a known-good committed fixture exposed the regression, then semantic checks explained it.
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.