Struggling with Recursive AI Agents? Harness2 Makes It Easy

Harness² gives developers a way to make AI agents better without touching the underlying model weights. The framework treats everything around the model – prompts, guardrails, memory, tool guidance, subagents, skills, scripts and middleware – as eight editable files. By refining these files the agent can improve its planning, tool use and domain knowledge while the model stays frozen.

The core idea is recursive agent harnessing. For each incoming task the agent first runs with the current harness (the reference). An editor then proposes two candidate harnesses: one that targets general agent failures such as poor planning or misuse of tools, and one that targets domain‑specific knowledge gaps. Both candidates execute the same task with the same model. Their trajectories are compared to the reference and only the edits that are supported by observed behavior are kept. This task‑level recursion needs no labels or test cases.

After each task a reflector stores the adopted harness together with an EXPERIENCE.md file that records the lessons learned. This library becomes the seed for future proposals, so the system improves across tasks through harness‑level recursion. The budget parameter k controls how many candidates are explored; parallel mode probes k invariant and k domain candidates per step, while sequential mode chains k refinement steps.

Harness² runs on Linux with Python 3.12 or 3.13 and uv, calling models through Vertex AI. No GPU requirement is disclosed. In experiments a single refinement step lifted GLM‑5 under OpenCode from 64.2 to 75.7 on WorkBuddy‑Bench (+11.5 points) and showed consistent gains across three benchmarks, while the worst case still improved by 1.2 points. The framework therefore offers a practical, plug‑and‑play path to better agents without costly retraining or label collection.

#AI #MachineLearning #LLMOps #Productivity #DeveloperTools #Innovation