Glossary

What is Outer loop?

The outer loop, in agentic engineering, is the cycle a human owns around an AI agent's work: verify the evidence that a change is correct, decide whether it ships, and carry the consequence once it does. Where the inner loop is capability (what the agent produces), the outer loop is accountability (what a person answers for), and it does not transfer to the agent just because the code arrived faster.

The distinction is set out in Own the Outer Loop, an essay by Addy Osmani, who led Chrome developer experience at Google for 14 years, and it has caught on fast among teams running agents in production. Stronger agents sharpen the point: the more of the inner loop they absorb, the more the decisions that remain (is this worth doing, does the evidence hold, do I approve it) concentrate in the loop a human still owns.

Owning the outer loop well takes infrastructure; willpower does not scale. Verifying a change means evidence you can trust rather than another model's opinion: the diff, the tests, and what historically breaks alongside it. Deciding means a hard gate, approve or block or redirect, before the change lands. Owning the consequence means a record that survives the run, so the decision and its justification can be answered for later. Skipping the outer loop is how a fleet ships code that no one understands, at a speed no one is checking.

How it relates to agent management

The outer loop is the accountability spine of AI agent management: its review, audit, and intervention disciplines are the machinery a human uses to verify, decide, and own what agents ship, at the pace a fleet sets. Earned autonomy widens the loop as an agent's record justifies it, without ever removing the human from it.