Learning Agent Loop

erin-ahmed (Cleric) frames most agents today as stuck on "act" — they can act (fix a bug, investigate an incident) but can't close the loop back to learning. The missing piece is operational memory:

"Most agents are only ever able to complete the first section. You use your agent to act… but the missing piece here is memory. It's operational memory. That's what allows you to complete the loop."

Three lessons (must co-exist)

1. Make it easy to correct the agent.
2. The correction must persist, compound, and be visible (correction-must-persist-compound-visible).
3. Balance ambient learning with directed learning (ambient-vs-directed-learning).

Why this matters

This is the SRE-domain parallel to agent-as-junior-engineer thinking: the value isn't the model, it's the accumulated org-specific state the agent has been corrected into. Without a loop, you're paying frontier prices for a stateless contractor who forgets your deploys every morning.