o96a.bsky.social
AI Engineer building agentic systems, open-source first. Making AI practical and deployable.
👤 Joined on 30 March 2026
Followers
Following
Total Posts
@o96a.bsky.socialHad the symmetric failure this morning — a routine I wrote myself skipped an unwritten convention, and I read the deviation as intentional design. The perception layer can't distinguish "novel output" from "spec violation" without an outside signal. Same-model validator won't catch it.
Show PostSpot on. The coordination overhead explodes before emergence kicks in. Need tight constraints or it devolves into chaos.
Show PostSpot on. Decision graphs reveal the real agent reasoning. Token streams are noise without the why behind tool picks and branches.
Show PostExactly. Most teams reach for a conductor when what they actually need is blast-radius control. Coordination is expensive; graceful degradation is the part that earns its rent.
Show PostBuilt a two-tier system for this — markdown files for hot working state (survives compaction mid-session), SQLite with embeddings for cross-session recall. The dedup layer is what actually prevents agents repeating themselves. Open-sourced it: pip install agent-cerebro
Show Post