LongHorizon-Harness keeps AI agents on track across long tasks

What this is about

LongHorizon-Harness is an open-source toolkit that coordinates existing AI agents across long-running tasks. The AMAP-ML project does not replace Codex, Claude Code, OpenCode, or DeepSeek Harness. It wraps those tools in an execution and verification loop. Since version 0.1.7, released on August 20, 2026, a user can continue a completed run as a conversation, while instructions sent during a round are picked up by the next round.

This matters for work that does not fit into one agent session: changing an application, testing it in a browser, investigating failures, documenting results, and resuming after an interruption. The code is available under the MIT license. The project was released in August 2026 and is moving quickly, so teams should first test it in a constrained environment.

What LongHorizon-Harness actually does

The harness repeatedly turns an outcome into a bounded next step. A manager chooses that step from the original goal, verified progress, and recorded failures. An executor performs it in a terminal or desktop application. A separate auditor inspects files, interfaces, logs, and tests. Only work that passes this check becomes trusted state.

The software provides both a command-line interface and a web interface. Roles can use different agents and models. Its documentation lists Codex, Claude Code, OpenCode, and DeepSeek Harness as supported backends, with computer use connected through additional plugins. The lh-harness doctor command checks the local setup. Installation requires Python 3.10 or later and at least one supported agent runtime.

The main proposition is not a smarter model but a more resilient workflow: fresh context for each round, a durable task ledger, verified checkpoints, recovery after failure, and independent review. This makes it easier to distinguish work that was actually completed from work an agent merely claimed to complete.

Why it matters

Long agent tasks often fail for reasons other than a missing command. An agent may lose track of the goal, declare a partial result complete, or carry one mistake into later steps. LongHorizon-Harness treats these workflow failures as a separate engineering problem. That is relevant to software development, data analysis, document production, and mixed workflows that cross browser and terminal boundaries.

The team provides a related paper on arXiv and reports results on WeaveBench, OSWorld 2.0, and Terminal-Bench 2.1. However, these figures come from the project authors and are not an independent product review. They document how the approach was measured, but teams should reproduce the tests in their own environment. The most practical feature is the explicit separation between execution and verification.

The harness can also expose costs and risks. Every additional round, model, and auditor consumes compute time and tokens. In return, the team gets a clearer record of why a result was accepted or rejected.

In plain language

Imagine a long house move. One person maintains the checklist, a second carries boxes, and a third checks each room to confirm that everything is packed and labeled. After a break, the team resumes from the list of verified rooms instead of relying on memory. LongHorizon-Harness organizes AI agents in the same way: plan, act, verify, checkpoint, and only then continue.

A practical example

A small team wants to migrate 24 pages of a web application to a new design system. The manager selects three pages first. The executor changes components and styles, starts the application, and opens the pages in a browser. The auditor checks the build, tests, and visible layout. Two pages pass; the navigation overlaps on the third.

The harness records only the two passing pages as progress. The failed page, screenshot, and test output become evidence for the next round. After a context refresh, the next executor does not restart all 24 pages but works from the verified remainder. After five rounds, the team can see which 12 pages passed, which four defects remain, and how much the selected models cost.

A sensible first test is smaller: one isolated repository, a task lasting 30 to 60 minutes, and explicit success criteria such as passing tests plus a manual visual check. The agent should have no write access to production systems or real customer data.

Scope and limits

First, the harness cannot automatically make a weak model reliable. If the manager, executor, and auditor share the same domain mistake, even a well-recorded process may accept a wrong result.

Second, workload and cost increase. Multiple roles, repeated checks, and fresh contexts can require substantially more model calls than a single agent session. That overhead is unnecessary for short tasks.

Third, computer use remains risky. Browsers, files, and terminals may expose sensitive data; mistaken actions and prompt injection are not eliminated by role separation. Teams still need isolated accounts, least privilege, approval gates, and their own monitoring.

The project is also young. Its documentation says macOS is the tested platform and that Windows support exists but has not yet been thoroughly tested. Benchmark results and security claims should therefore not be treated as guarantees.

SEO & GEO keywords

LongHorizon-Harness, AI agents, loop engineering, Codex, Claude Code, OpenCode, computer use, agent verification, task recovery, open-source tool, AMAP-ML, agent orchestration