Machines of Code and Grace

They explore using agents with Obsidian to manage personal organization and local integrations. They introduce Change Intent Records as tiny decision logs to record why choices were made. They debate prompts as the evolving source of truth for intent. They wrestle with the verification gap and ways agents might use screenshots, video, or alternate checks instead of human eyeballs.

A playful dive into attention checks like Van Halen’s brown M&M clause as a canary for missed details. They explore simple telltales such as printing a trumpet to confirm an agent followed instructions. Conversation covers deterministic checks, evals as golden masters, and using stubs or guardrails to surface compatibility issues.

A lively chat about whether model context size changes agent behavior and project fit. They compare model roles and tradeoffs between cheaper, faster models and heavier ones that handle multi-file changes. They dig into why human review struggles with long project histories and propose agent roles, PR annotations, and layered AI systems to preserve domain knowledge.

They explore defining AI feature owners to prevent accidental breakage. They use a card-shuffler example to show how overlooked interactions cause bugs. They explain making agents that own knowledge, invoke reviewers during planning, and update knowledge bases. They discuss dividing attention with sub-agents and documenting new feature owners to keep systems reliable.