Google researchers open-source Harness², which tunes an agent's harness instead of its model

A harness is everything that turns a model into an agent: prompts, tools, memory, skills and orchestration[1]. Harness² exposes eight editable parts of it, namely the system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins[1].

The authors' starting point is that the same model can perform very differently depending on that scaffolding, and that hand-tuning it for each new kind of task doesn't scale[2].

Harness² runs two loops. The first improves the harness for the task in front of it with a Propose–Probe–Compose step: propose edits, run the same task with each, and compose a refined harness from the differences between those runs, without needing a known correct answer[1]. The second loop saves the adopted harnesses and evidence-backed lessons into a library that shapes how later tasks get refined[1]. It can scale by trying several candidates in parallel or by refining one after another[1].

The team tested it on WorkBuddy-Bench (office, web and code work), JobBench and Harvey's Legal Agent Benchmark, using frozen Gemini 3.5 Flash and GLM-5 models inside the off-the-shelf OpenCode and Codex harnesses[2]. A single improvement step raised scores by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on the legal benchmark's partial pass rate[2].

Starting from OpenCode with five parallel candidates, Harness² beat the best unmodified production harness on each WorkBuddy-Bench domain by 4.6 to 5.8 points[2]. Claude Code was the strongest fixed harness on code and office tasks in that comparison, and Codex on web tasks[2]. Notably, the default harness editor in the experiments was Anthropic's Claude Opus 4.8[2].

The extra performance isn't free: proposing and probing changes means more executions and more inference cost[2]. The authors also say comparing runs doesn't guarantee an edit is correct or will carry over to a different task, and their failure analysis includes a case where the system picked an output scoring 55 over one scoring 95[2]. The repository says it is not an officially supported Google product[1], and the setup guide warns that the Codex executor runs with its sandbox disabled[3].

If you run agents on repeatable professional work, this is a concrete argument that tuning your harness can buy as much as a model upgrade. You don't have to adopt Harness² wholesale to use the idea: test a prompt, skill or tool change against the same task, keep what the runs support, and write down why. If you do try the code, run it in an isolated environment with network limits, as the setup guide suggests.