Possible complementary direction: LongHorizon-Harness for sustained agent tasks

[Upper9527]

Hi, I'd like to suggest a complementary direction for better-harness.

Project:
https://github.com/AMAP-ML/LongHorizon-Harness

Your project focuses on coding-agent and software-engineering workflows.

LongHorizon-Harness focuses on carrying one task across many rounds without state drift. It keeps the goal and verified progress in durable state, starts each execution round with a fresh context, and uses an independent Auditor before accepting progress.

Why this may be useful here:

• A small integration note could clarify how the existing execution layer can sit inside a recoverable, independently checked long-running task.
• it adds an outer task-lifecycle layer without claiming to replace the project's existing agent or runtime

If this direction fits, I can prepare a small documentation or example contribution following the repository's conventions.

I am part of the team behind LongHorizon-Harness, so I wanted to disclose that connection.

Paper:
https://arxiv.org/abs/2608.01964

Project page:
https://lh-harness.pages.dev

[phodal]

Thanks — and thanks for disclosing your affiliation up front. The complementary read makes sense; Better Harness is a review/evidence layer, not a runtime.

Feel free to give it a try: open a spec under docs/specs/-.md (see AGENTS.md) covering where the boundary sits and how a completed run could be read as session evidence. Worth noting I plan to add runtime capabilities here too, so treat the spec as a proposal to discuss rather than a merge commitment.