Harness Engineering Resources

6 min readUpdated Jul 14, 2026

Created: 2026-06-09 11:15
#note

A curated landscape of open-source projects, guides, and papers in the field of Harness Engineering. It complements the conceptual hub and the practical Building an Agent Harness from Scratch note by pointing to concrete implementations one can read, run, or fork. Entries are grouped by what they are — a runtime to build on, a guide to learn from, or a research artefact — because the right resource depends on whether the goal is to ship, to learn, or to study the frontier.

Harness Runtimes (build on these)

These are working agent runtimes — the substrate a harness rides on. They differ mainly in language, provider coverage, and how opinionated they are.

Pi (earendil-works/pi) — a TypeScript monorepo billed explicitly as an "agent harness": its own agent runtime (pi-agent-core), a unified multi-provider LLM API (pi-ai), a coding-agent CLI, and a terminal UI. Notably it ships no built-in permission system — the project tells users to sandbox or containerise it themselves, a deliberate "thin harness" stance. A clean example of a harness with its own loop that treats frontier models as interchangeable providers rather than building on any vendor SDK.
goose (aaif-goose/goose, formerly block/goose) — a mature, general-purpose agent written in Rust, now stewarded by the Linux Foundation's Agentic AI Foundation. Desktop app, CLI, and embeddable API; 15+ providers; tools via the MCP Protocol; skills and lifecycle hooks. The most production-hardened of the open runtimes.
Deep Agents (langchain-ai/deepagents) — "the batteries-included agent harness," built on LangGraph, aimed at long-running tasks with planning, file tools, shell, and sub-agents out of the box. Includes a Ralph-mode example (a hook that forces the agent to keep working on exit). This is the codebase behind LangChain's harness-engineering experiments described in Harness Middleware Techniques.
OpenHarness (HKUDS/OpenHarness) — a Python implementation with a query→stream→tool-call loop, 40+ tools, on-demand skill loading, safety permissions, and lifecycle hooks; ships a personal agent ("ohmo") that opens its own pull requests.
HarnessX (Darwin-Agent/HarnessX) — a Python "harness foundry" that treats the harness as a composable, first-class object: agent = model.agentic(harness), any behaviour expressed as a Processor piped across a 9-dimension pipeline. Unusually it ships two evolution loops (harness search + RL model training) out of the box, so it doubles as an entry in the automated-optimisation section below. Developed in its own note: HarnessX (Harness Foundry).

Meta and Skill-Based Harnesses

Harnesses expressed as composable skills, or tools that generate harnesses.

Atelier (martinffx/atelier) — the skill-based Claude Code harness already covered in Atelier (Agent Harness).
revfactory/harness — a meta-skill that designs domain-specific agent teams and generates the skills they use; harness-building as a skill rather than a framework.
Chachamaru127/claude-code-harness — a Plan → Work → Review autonomous cycle layered over a coding agent; a concrete instance of the Research-Plan-Implement Loop expressed as verb-skills.

Guides and Learning Resources

Harness Engineering Guide (nexu-io/harness-engineering-guide, harness-guide.com) — an open guide with runnable code covering the agentic loop, tool systems, memory and context, guardrails, sandboxing, sub-agents, and multi-agent orchestration. A good first-principles reference; several of its chapters surface concepts not yet captured in this vault (see "Concepts worth extracting" below).
learn-harness-engineering (walkinglabs) and learn-claude-code (shareAI-lab) — zero-to-one tutorials that build a harness incrementally, each lesson motivated by an observed failure.
Awesome lists — ai-boost/awesome-harness-engineering, walkinglabs/awesome-harness-engineering, Picrew/awesome-agent-harness, and the weekly-ranked RyanAlberts/best-of-Agent-Harnesses (100+ harnesses scored). YennNing/Awesome-Code-as-Agent-Harness-Papers collects the academic side. GitHub topics [harness-engineering] and [agent-harness] are the live feed.

Research: Automated Harness Optimisation

The frontier is treating the harness itself as the object of search, with the model held fixed. A useful way to organise the entries is who does the improving: a human, a stronger external agent, or the agent itself. This vault already covers AgentFlow; its companions:

Meta-Harness (Stanford IRIS Lab; arXiv:2603.28052) — searches over harness code with an agentic proposer that reads the source, scores, and full execution traces of all prior candidates. A stronger external agent improving a target harness. Developed in its own note: Meta-Harness - End-to-End Optimization of Model Harnesses.
Self-Harness (Shanghai AI Lab; arXiv:2606.09498) — the agent improves its own harness with no human and no stronger teacher, via a Weakness Mining -> Proposal -> Validation loop; up to +21.4pp / +138% on Terminal-Bench-2.0 with the model frozen, and the retained edits are model-specific. This is the third paradigm, self-improvement, completing the human / external-agent / self triad. Developed in its own note: Self-Harness - Harnesses That Improve Themselves.
Agentic Harness Engineering / AHE (china-qijizhifeng/agentic-harness-engineering) — concurrent work on observability-driven automatic evolution of a coding-agent harness; reports a frozen, discovered harness that transfers across base models and from Terminal-Bench 2 to SWE-bench-Verified without re-tuning. The transfer result is the interesting claim: a harness as a reusable artefact rather than a per-target build.
HarnessX (Darwin-Agent/HarnessX) — the runtime that operationalises the above: an in-repo MetaHarness runs the Self-Harness-style self-observation loop, and reward-annotated trajectories feed RL model training, so harness and model can co-evolve. Listed as a runtime above; note at HarnessX (Harness Foundry).
TerminalBench-2: the long-horizon, dependency-heavy command-line benchmark most of the above optimise against, and the de facto yardstick for coding-harness quality.

Concepts Worth Extracting

Beyond the named papers, the guides above surface several transferable ideas not yet in this vault — candidates for their own notes: classifier-based permissioning (replacing approval fatigue with a model-based gate), eval-infrastructure noise (resource configuration alone swinging benchmark scores by several points), long-running harness design (context anxiety, self-evaluation bias, a generator-evaluator split), the brain/hands/session decoupling of managed agents, and large parallel agent teams coordinating through git. These extend Harness Middleware Techniques and the Managed Agent Harness (Bedrock AgentCore) note.

References

Harness Engineering, Building an Agent Harness from Scratch, Harness Middleware Techniques, Atelier (Agent Harness), HarnessX (Harness Foundry), Managed Agent Harness (Bedrock AgentCore), Research-Plan-Implement Loop, Loop Engineering, Synthesizing Multi-Agent Harnesses for Vulnerability Discovery, Meta-Harness - End-to-End Optimization of Model Harnesses, Self-Harness - Harnesses That Improve Themselves, Agentic AI Frameworks, MCP Protocol