The gist

LongHorizon-Harness is an open-source framework from AMAP-ML for managing autonomous AI agents performing complex, long-running tasks. It doesn't replace an agent but acts as an execution, state-management, and verification system. It enables agents to operate for extended periods on real computers, working across both GUIs and the command line, by ensuring that progress is continuously verified and preserved, preventing state drift on multi-hour tasks.

What it does

Manages complex tasks by separating them into Manager, Executor, and Auditor roles.
Persists verified progress, allowing agents to continue from a trusted state even if an action fails.
Operates across both desktop GUI applications and the command-line interface within a single task.
Integrates with multiple agent backends like Claude Code, Codex, and OpenClaw.
Provides a dashboard for real-time visibility into task plans, execution results, and audit evidence.
Stores a complete, inspectable audit trail for every task run, ensuring reproducibility.

How it works

LongHorizon-Harness is a Python CLI tool installed locally. A user defines a complex task via a command-line argument. The harness then directs a connected AI agent (e.g., GPT via Codex) to execute the task in rounds. It uses separate roles for planning, execution, and auditing; only results that pass independent verification are saved to the persistent state. The system is open-source and works with local, SSH, or Docker execution environments.

Best for

Developers and researchers building or evaluating autonomous agents for complex, multi-step computer tasks where verifiable progress and resilience to failure are critical.

Watch out for

This is not a standalone agent or model. It is a harness that requires and runs on top of an existing agent backend, such as Claude Code, Codex, or OpenClaw.