A 4-loop framework for production AI agents focuses on task, verification, and improvement

MaryamMiradi · x · 2026-07-29

Maryam Miradi lays out a 4-loop framework for production AI agents: a task loop, a verification loop, an event-driven run loop, and a system-improvement loop.

The four loops

Loop 1: Do the task — the model plans, uses tools, observes results, and stops when it thinks the job is done.
Loop 2: Check the result — don’t trust the agent’s self-assessment; validate against tests, rules, rubrics, or evals.
Loop 3: Run when work arrives — trigger the agent from Slack, GitHub, email, queues, or schedules, then send the result back where humans already work.
Loop 4: Improve the system — analyze traces across many runs, find recurring tool failures and missing context, then update prompts, tools, skills, evals, or workflow.

The key message is that human approval can sit inside any loop: before risky tool calls, before accepting the output, before merging, or before changing the agent itself.

Original post →

A 4-loop framework for production AI agents focuses on task, verification, and improvement

More from coding & agent