HeyHomees Lab Notes · No. 4 · Google DeepMind · Published October 19, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Harness² adds a recursion loop that proposes, probes, and composes harness edits without ground-truth feedback.
- A single recursion step improves GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 % on LAB.
- Runs on Linux with Python 3.12, Node.js 20+, and optional Google Cloud CLI; no model credentials needed for the local demo.
What is it?
Harness² is a plug-and-play framework for recursive agent harnessing in an open world, meaning agents can encounter tasks unseen during harness design. It exposes eight editable components such as system prompt, guardrails, memory, and tools, and supports both parallel and sequential recursion modes.
How does it work?
The system runs a contrastive Propose–Probe–Compose cycle: it proposes harness edits, probes their effects on the current task, and composes a refined harness from behavioral contrasts, all without ground-truth feedback. Task-level recursion refines the execution harness per task, while harnessing-level recursion retains adopted harnesses and lessons for future tasks.
How well does it do?
Across three professional benchmarks and four model-harness configurations, a single Harness² step yields gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB's partial pass rate. Increasing the task-level recursion budget provides further gains.
What does it need to run?
It requires Linux, Python 3.12, and the uv package. Model-backed runs also need Node.js 20+, the Google Cloud CLI, and benchmark-specific dependencies (e.g., Pandoc and Poppler for LAB, Docker for WorkBuddy-Bench). No model credentials are needed for the local demo.
By the numbers
- WorkBuddy-Bench gain: 11.5 points
- Task-level recursion budget option: --k 1
- Parallel recursion option: --k K --mode parallel
- Sequential recursion option: --k K --mode sequential
The honest read
The reported gains (up to 11.5 points, 4.5 points, and 11.8 % respectively) are self-reported by the Google DeepMind team on three specific professional benchmarks; they do not reflect comparisons to other frameworks or to non-recursive baselines. The improvements come from a single recursion step, and larger gains are only hinted at when the task-level recursion budget is increased, but exact numbers for higher budgets are not provided. No ground-truth feedback is used, so the scores measure performance under the authors' evaluation pipelines rather than absolute task mastery. The framework is open-source and runs without model credentials for the demo, but model-backed evaluations require external APIs and incur costs, which are not quantified in the source. The results therefore apply to the specific models (GLM-5 and Gemini 3.5 Flash) and configurations tested, not to all possible agents.
What we don't know yet
- Exact performance when using multiple recursion steps beyond the single step reported.
- How Harness² compares to other agent-orchestration frameworks on the same benchmarks.
- Resource usage (CPU/GPU, memory) for large-scale runs.
Sources
- README, GitHub (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.