Ambient vs Directed Learning

Framing from erin-ahmed (Cleric) on how agents should acquire new knowledge:

"If you make it easy to correct but you don't ambiently learn then you're only ever learning from where the user directs you. It limits your learning. And if you're ambiently learning but you don't allow the user to correct you then you run the risk of compounding errors… think about all the dead code or the outdated documentation that's just laying in wait for the agent to stumble on."

• Directed learning — user explicitly teaches the agent a rule/procedure.
• Ambient learning — agent infers rules by watching how the team operates, reading code, observing incident channels.

Why this matters

Pure ambient learning ingests every stale artifact in your org (dead code, outdated runbooks) and compounds the rot. Pure directed learning caps at your willingness to teach and never scales. The durable architecture does both and exposes the ambient inferences for a human to confirm or reject — closing the loop between passive absorption and explicit reinforcement.