Three Loops in an Agent System
What people call the agent loop is only the innermost of three loops. The outer loop finishes tasks, and the meta loop is how an agent gets better over time.
A few weeks after North went live, I had a question. It worked, and it sent its reminders. But did it get better over time?
The answer was no, not unless I designed how it gets better. What people call “the agent loop” is only the innermost of three loops.
Three loops
The talk by 「大大帶我飛」 splits an agent system into three loops:
- With only the inner loop, the agent acts, but you do not know whether it finished.
- With the outer loop too, the agent can finish tasks, but it stays just as good, and just as bad, forever.
North’s three loops
- Inner: in one reminder, North reads files, runs scripts, and writes the message.
- Outer: one scheduled trigger sends one reminder. When something must change, it stops and waits for my yes (see Few Tools, Clear Boundaries).
- Meta: every judgment goes into a decision log. Once a month, a scan looks for judgments that repeat.
The meta loop needs two things. The first is a record, which I cover in How Do You Know the Agent Did Well?. The second is a way to turn the record into improvements, which I cover in Graduate NFR into FR.
An agent that works needs one loop. An agent that gets better needs three.
References:
- A talk by 「大大帶我飛」
Related: see the record the meta loop needs in How Do You Know the Agent Did Well?, how that record becomes improvements in Graduate NFR into FR, or go back to the series overview: Design an Agent.