Back to Project
H

Harness2

Project Oct 19, 2026

Research framework that refines agent execution harnesses through propose-probe-compose steps and retains lessons across tasks while keeping model weights fixed.

Visit primary link 3 sources attached
Harness2 authors' pipeline diagram connects agent execution, proposed harness edits, probes, composition and a persistent improvement harness for later tasks.
Author-published research workflow diagram from the reviewed Harness2 revision, with the full frame preserved. It explains the proposed two-level recursion, not a live run, independently reproduced result or proof that benchmark access is controlled. Source: Harness2 authors / google-research, Apache-2.0 repository
agent-ecosystemai-agentsagent-security

What happened lately

Harness2 timeline

Harness2, styled Harness², is an Apache-2.0 research framework in the google-research repository. It changes the prompts, tools, memory and other components around an agent’s model, with task-level experimentation and persistent lessons for subsequent tasks. It does not update model weights.

The reviewed release supports OpenCode and Codex executors and the JobBench, LAB and WorkBuddy-Bench benchmarks. Linux and Python 3.12 are the documented starting setup. A deterministic local demo needs no model credentials; model-backed runs require configured providers and incur API costs.

The project documents unrestricted executor network access and disabled Codex sandboxing, including an observed validation rollout retrieving a public grading rubric. Benchmark experiments need appropriate network controls. Its authors explicitly disclaim official Google product support.