Glossary

What is Inner loop?

The inner loop, in agentic engineering, is the cycle an AI agent runs on its own to produce a change: investigate the problem, implement a solution, test it, and report back. It is where an agent spends most of its effort, and the part of software work that recent models and harnesses have most substantially automated.

The phrase is borrowed from developer experience, where the inner loop long meant the fast edit-build-test cycle a single engineer repeats all day, as opposed to the slower outer loop of review, merge, and release. Applied to agents, the inner loop is everything that happens between your prompt and the agent's proposed change: reading the code, forming a plan, writing the diff, running the tests, and iterating until it looks done.

The inner loop is capability. What it deliberately excludes is the judgment about whether the change should exist, whether its evidence actually holds, and who answers for it once it ships. The framing comes from Own the Outer Loop, Addy Osmani's essay: the more an agent can do unattended, the more the decisions that remain concentrate in the outer loop a human still owns.

How it relates to agent management

The inner loop is the capability half of AI agent management: the work an agent performs. Management exists to wrap it in the outer loop of review, records, and accountability that the inner loop does not supply for itself.