$ cd homeguild/labs
Architecture decisions, agent design, debugging sessions, and patterns from building an AI-first field service platform.
When a conversational AI with log access and a coding agent with codebase access share reference points but observe the system independently, you get something more reliable than either alone. This post explores the independent verification layer that emerged naturally during HomeGuild's production debugging sessions.
Why we threw out the original Atticus implementation and how the new LangGraph architecture with a user profile, business outline, and evaluator doc produces fundamentally better business plans.
Named after the Roman night watch. How a cron-driven Postgres query and a stateless worker queue creates proactive business awareness without burning tokens on the scheduler itself.
The Android build worked fine via Bearer token. The Flutter web build silently set hg_token to blank, causing null auth on every GraphQL request. A walkthrough of the investigation and fix.
Why we maintain a CLAUDE.md per agent directory and what belongs in it — architectural constraints, naming conventions, prior decisions, and things the agent should never do.
How we removed static credentials from the pipeline entirely using --dart-define-from-file injection and Firebase App Check with Play Integrity.