Plan once, then hand off the loop
You define the finish line once, then AgentLoop routes work through a planner, fresh worker, and independent critic instead of asking you to relay every message.
AgentLoop
AgentLoop is a local coding loop tool that sends work to fresh workers, reviews each cycle with an independent critic, and repeats until your project meets your standards.
Overview
AgentLoop is a local coding loop tool that runs unattended on your machine. It is built around a simple handoff model: you set the goal, define the standards, and AgentLoop sends work to fresh Codex workers, reviews the results with an independent critic, and repeats until the project meets your rubric.
The site positions the product for tasks where you would otherwise babysit the work: feature builds, migrations, hardening passes, and product polish. Rather than relying on one long-running context, it uses fresh processes and project files to carry evidence forward between cycles.
Core capabilities
Fresh workers for each cycle
Each cycle starts with a clean worker process so the next attempt does not inherit stale context from the previous one.
Independent critique step
A separate critic evaluates the result using your rubric, so success is based on the product's standards rather than only on test status.
Files as memory and evidence
Project files serve as the working memory between cycles, with evidence carried forward through files such as GUIDELINES.md and cycle logs.
Local-first execution
The product runs on your machine where the code lives, with local execution and no hosted workspace required.
Cycle observability and control
The dashboard and replay views expose cycle-by-cycle output, verdicts, transcripts, costs, and file handoffs for review and control.
Where it fits
Feature development
Use AgentLoop when you have a well-scoped feature and acceptance criteria, and you want the tool to carry the implementation through several edit-and-review cycles.
Code migration
Use it for migration work that needs repeated verification, especially when the changes touch multiple files or require regression coverage.
Bug fixing and hardening
Use it for hardening passes where you want an independent critic to catch issues that tests might miss and then ask for a follow-up fix.
Polish and cleanup
Use it for product polish tasks where the code needs to be brought up to a defined standard rather than simply made to pass tests.
Reviewable unattended runs
Use it when you want unattended progress but still need visibility into what changed, why a critic approved it, and what evidence was carried forward.
Pros and Cons
Pros
- Runs against a local codebase instead of a hosted workspace.
- Uses fresh workers and a separate critic to reduce reliance on a single conversation context.
- Lets project files and GUIDELINES.md carry standards and evidence across cycles.
- Provides replay, logs, verdicts, and cost visibility while a run is in progress.
- Can be cancelled from the dashboard or checked through ChatGPT via MCP.
Cons
- The source does not show pricing, plan structure, or usage limits.
- The current page suggests Codex CLI is required, so setup depends on that external tool.
- The site mentions future support for more engines and two-way agent questions, which implies those capabilities are not yet available.
FAQ
What do I need to run AgentLoop locally?
AgentLoop is designed to run against a local project and requires Node.js 18+, Git, and an authenticated Codex CLI. The source says there is no npm install and no hosted workspace.
How does the workflow work?
The page describes a loop where you define the goal and standards in ChatGPT and GUIDELINES.md, AgentLoop sends work to fresh workers, an independent critic reviews each result, and the next cycle carries the evidence forward.
What kinds of coding tasks is it meant for?
The source shows it can be used for feature work, migrations, hardening passes, and product polish, especially when you want a bounded goal and clear acceptance criteria.
Can I monitor or stop a run while it is working?
The page says the critic runs in a separate fresh process, passing tests are only evidence, and the system can watch every cycle, verdict, transcript, cost, and file handoff. It also mentions cancelling from the dashboard or asking ChatGPT for status through MCP.
Does AgentLoop support other engines or integrations?
The site states that Codex ships first and that the engine boundary stays pluggable. It also says MCP keeps planning separate from local execution, with future support mentioned for more engines, two-way agent questions, and research loops.
Quick Facts
- Category
- Developer Tool
- Platform
- Local machine; Windows, macOS, and Linux are shown
- Workflow
- Plan in ChatGPT, run local Codex workers, review with an independent critic
- Requirements
- Node.js 18+, Git, and authenticated Codex CLI
- License
- MIT licensed
- Source domain
- agentloop.tools
AgentLoop Alternatives
ByteAsk
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repos and verifies changes with compilers, debuggers, sanitizers, and tests.
Ghost
Ghost is a terminal-based AI assistant for chatting, code generation, and CLI tasks. Includes free models, supports Linux, macOS, Windows, and is open source.
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads in Firecracker micro-VMs with private networking and SDK, CLI, or MCP control.
hob
hob is an independent workspace for coding agents, with local control over sessions, terminals, history, routing, and follow-up work.
Manta AI
Manta AI is an autonomous web app testing tool that maps app behavior, catches regressions, and generates tests from a URL, no scripts or selectors needed.
SonOf
SonOf connects to your repo and PM tool, audits your codebase, and turns approved work into shipped tickets with senior engineering review.