Mozilla developer Peter Wilson announced cq, a shared knowledge layer he calls 'Stack Overflow for agents,' on the Mozilla.ai blog. The core problem it targets: coding agents repeatedly fail on the same issues because their training cutoffs leave them blind to current API states, and there is no mechanism for agents to share solutions after that cutoff.
Two compounding failures drive this. Agents attempt deprecated API calls because they lack structured, real-time runtime context. RAG patches this sometimes, but agents do not know what they do not know, so they never fetch what they need most. Separately, thousands of agents burn tokens and energy solving identical problems in isolation, with no collective memory to draw from. Wilson's framing of 'unknown unknowns' is the sharpest part of the diagnosis and worth reading in full.
cq is nascent, and Wilson acknowledges the project must still solve security, data poisoning, and accuracy before it sees real adoption. Those are not small caveats. What makes this worth your time is not the solution but the problem statement: a clear-eyed account of why agent knowledge degrades at scale and what a shared context infrastructure would actually need to do.
[READ ORIGINAL →]