Harness2

Harness2 is an open-source agent framework that recursively optimizes prompts, tools, and memory for open-world tasks.

Harness2 is a plug-and-play framework for recursive agent optimization in an open world where agents encounter unseen tasks. It connects two levels of recursion while keeping model weights fixed, refining execution structures for current tasks through a contrastive propose, probe, and compose step. The system updates a persistent improvement structure with adopted components and lessons from each task to guide future refinements. It exposes eight editable components around the model including system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. The release supports parallel and sequential recursion, OpenCode and Codex executors, and JobBench, LAB, and WorkBuddy-Bench benchmarks.

Harness2 at a glance

Best for: Software engineer
Runs on: Linux, Python, CLI, open source
Pricing: Free and open source under Apache 2.0 license
Released: October 2026
Category: AI agent tool
Website: github.com/google-research/harness2

Key features

Task-level recursion with contrastive propose, probe, and compose steps
Harnessing-level recursion updating persistent improvement structures
Support for parallel and sequential recursion modes
Eight editable components including prompts, guardrails, and memory
Integration with OpenCode and Codex executors
Compatibility with JobBench, LAB, and WorkBuddy-Bench

Use it for

Optimizing agent execution structures for unseen professional tasks
Running legal work benchmarks via LAB
Evaluating office, web, and code tasks using WorkBuddy-Bench
Executing JobBench professional work scenarios

Alternatives to Harness2

HarnessRouter — One API to manage agent runtimes, sandboxes, and execution lifecycles.
OpenViking — Unifies agent memory, knowledge RAG, and skills into a single directory structure.
Harness Meta-Skill — Generates specialized agent teams and skills for Claude Code from a single prompt.
Harness Engineering — Improves agent output by shaping the environment with context and tools.
OmniAgent — It evolves skills in real-time and dynamically hardens security during task execution.
Exo Harness — Separates trusted state and security from agent logic for safe execution.
Composio — Equips AI agents with tools, managed user auth, and sandboxes across 1,500+ apps
LongHorizon Harness — Maintains durable task state and executes reliable, verifiable workflows over many hours.
Webwright — Enables code-based web automation with reusable and rerunnable scripts.
Rax — Exploration compiles into code that runs up to 241 times faster.
OpenSquilla — Maximizes AI intelligence density while minimizing token costs through smart routing.
Raven — Provides a persistent, self-evolving environment for all-domain agent collaboration.

All Harness2 alternatives