[process] | 2026-06-12 | 8 min

# Inner Loop / Middle Loop / Outer Loop: The Three Closed Loops of Agent Work

Agent work isn't linear — it splits into three nested closed loops: the inner loop (code-test), the middle loop (task-review), and the outer loop (metrics-improvement). Understanding this model is foundational to designing AI4SE processes.

[workflow-design][process]

The Three-Loop Model

Outer Loop (weeks/months): Measure → Improve Rules/Process/Methods
  └── Middle Loop (hours): Plan → Execute → Review
        └── Inner Loop (seconds/minutes): Code → Test → Fix

Key Activities per Loop

Loop	Cadence	Driven by	Output
Inner	Seconds to minutes	Agent	Code + tests + immediate fixes
Middle	Hours to days	Human + Agent	Completed tasks + passing review
Outer	Weeks to months	Human-led	Metrics → updated Rules / Spec / Harness

Why the Three Loops Matter

Different loops need different metrics
Inner optimizes for speed, Middle optimizes for quality, Outer optimizes for the system
Optimizing only the Inner loop = a local-optimum trap

In Practice

Inner: leave it to the Agent, intervene minimally
Middle: place gates at Review and Plan
Outer: check metrics periodically, adjust global constraints