Concept 04

Agentic loops

An agent is a model running in a loop: it plans, takes a step, looks at what happened, and decides the next step, until the job is done or it is stuck. The loop is what lets an assistant fix its own mistakes, retry with a different approach and finish multi-step work without you supervising each step. Some practitioners call shaping this loop loop engineering.

Standing: Established since 2025-04
Stars, week: +1,238 across 5 repositories
Stars, month: +7,427, through 2026-11-09
Listings: 5 chosen by hand

The standing

Where it stands

Among the 14 concepts. Bars: GitHub stars gained this week.

Established · 7

Rising · 7

How to use it today

In ChatGPT or Claude, as a person doing the work, not as a developer.

1Use the agent side of your assistant (ChatGPT Work, Claude, Claude Code, Codex) for work with several steps and a result you can check, not for a one shot answer.
2State the goal and the stopping condition ("a table of the ten competitors with pricing, sourced; stop when every row has a source"), not the steps.
3Ask it to plan first and show the plan; approve or edit it before it acts.
4Give it a way to check its own work (a test, a checklist, a target format) so the loop has something to verify against.
5Set a budget: how long it may run, which actions need your confirmation.

What changes day to day

You move from editing answers to reviewing plans and results. Longer tasks run while you do something else; your attention goes to the goal statement and the acceptance check, which become the important writing.

Where it goes wrong

A loop without a verification step runs confidently in the wrong direction.

Open-ended goals ("improve this") never terminate well; give it something to satisfy.

Listings that use it well

Chosen by hand after reading them, each with the reason. Install from the listing's own page.

Listing	Kind	Why it is here	Stars
ai-agent-book	Agent skills	Skills on loop engineering and error recovery from the agent design book.	58k
engineering	Agent skills	agent-harness compiles a goal into a verifiable plan, executes, verifies with machine checks, retries with caps and escalates; zero-hallucination-coder runs Discuss, Map, Decompose, Execute, Verify.	30k
oh-my-claudecode	Agent skills	ralph runs a self referential loop until completion with a verification reviewer; autoresearch is a stateful improvement loop with an evaluator contract.	41k
loop-engineering	Agent skills	Patterns, starters and CLI tools for loop engineering, with loop-verifier as a separate rejecting reviewer and loop-budget for early exit.	12k
ralph-tui	Agent skills	Converts PRDs into task files for the Ralph autonomous loop runner.	2.5k

Elsewhere

Repositories that embody the idea and are not in the catalog.

openai/codex ↗OpenAI's open-source terminal coding agent; the loop reads files, runs commands and iterates.
SWE-agent/SWE-agent ↗The academic agent-computer-interface loop that first posted real SWE-bench numbers (NeurIPS 2024).
OpenHands/OpenHands ↗Open platform whose event-driven loop records every action and observation for replay.

Standing

Our judgment, with the reason and the sources it rests on.

Established since 2025-04. Every coding assistant and both main chat apps now run this loop: plan, act, check, repeat. Copilot agent mode (February 2025), Claude Code (February 2025) and Codex CLI (April 2025) brought it to coding; ChatGPT Work and Claude brought it to office work in 2026. The basic pattern is settled; how to design the checks and stopping rules around it is still moving, under names like harness engineering and loop engineering. Settled practice: the methods are known, the tools are mature, and the question is how well you use it.

Claude 3.7 Sonnet and Claude Code2025-02-24
Vibe coding with GitHub Copilot: Agent mode and MCP support rolling out to all VS Code users2025-04-04
GitHub Copilot: The agent awakens2025-02-06