On 2026‑10‑18 I announced that Google Research open‑sourced a recursive Agent Harness called Harness². It is fully open‑source and not an officially supported Google product. The harness is defined as everything that turns a model into an agent – system prompt, tools, memory, skills, sub‑agents, and guardrails.
Harness².name = Harness²
Harness².origin = Google Research
Harness².status = 100% Open Source
Harness².repo_url = https://github.com/google/harness2
Harness².definition = everything that turns a model into an agent: the system prompt, tools, memory, skills, subagents, and guardrails
Harness².workflow = Propose edits across eight parts of the harness, from the system prompt to skills, scripts, and plugins
Harness².probing_mode = parallel and sequential
Harness².self_editing_decision = keep model weights frozen and edit only the harness
Harness².improvement_score = up to 11.8 points after one round of self‑editing
Harness².local_demo_available = true