The layer between the model and the work
An agent harness is every piece of code, configuration, and execution logic that is not the model itself: the loop that keeps calling the model, the tools it may run, the sandbox those tools run in, the memory that survives a restart, and the rules that decide what reaches the context window. In 2026 this layer got a name, a plugin standard, and nine products worth choosing between. HarnessKit helps you pick one - and set it up so the choice survives the next model release.
Run the 60-second chooser Get the starter kit
Reasons. Claude, GPT, Qwen, DeepSeek, open weights - the part everyone argues about, and increasingly the part that is much of a muchness at a given tier.
Runs one agent: the loop, tools, sandbox, memory, context rules. Given models of similar capability, the harness is often the distinguishing factor in whether the agent actually works.
Composes several agents with explicit state and branching. The rule-based engine of the stack: you write the branches, you get the determinism.
Runs many harnesses across a team over time: durable execution, cost attribution, governance. This is where organizations end up; few need to start here.
Layer framing per LangChain, MongoDB, and the winder.ai comparison. Pydantic AI's v2 pull of harness features into its framework core is the clearest evidence the harness layer is real.
| Harness | Maker | License | Models | Stars (Aug 2026) |
|---|---|---|---|---|
| Claude Code | Anthropic | Proprietary (custom license + commercial terms) | Anthropic models only | 142k |
| Codex | OpenAI | Apache-2.0 | OpenAI models only | 107k |
| OpenCode | Anomaly | MIT | 75+ providers via Models.dev, including local models | 199k |
| Qwen Code | QwenLM (Alibaba) | Apache-2.0 | OpenAI, Anthropic, Gemini and Qwen APIs plus local Ollama/vLLM | 27.2k |
| DeepSeek Harness | DeepSeek | MIT | Any - everything is a plugin, including the agent loop | 95k in ~2 days after launch |
| Goose | Agentic AI Foundation (Linux Foundation) | Apache-2.0 | 15+ providers including local Ollama | 53k |
| Zed Agent | Zed Industries | GPL-3.0-or-later | Drives Claude, Codex or OpenCode over the Agent Client Protocol | 88.9k |
| OpenHands | All-Hands-AI | MIT | Any model; ACP client | 85k |
| Pydantic AI Harness | Pydantic | Open source | Any Pydantic AI supports | n/a (library) |
Feature counts are the wrong axis - the features are converging (three of the nine already speak the Agent Client Protocol and can drive each other). What separates them is what each one assumes you have already decided: the model, where it runs, and who gets to change the agent's behavior. The chooser works on those assumptions.
Critic Han Lee calls harness work hidden technical debt: most of it dissolves into the next generation of models. The test for any piece of harness work is whether a better model would make it unnecessary. Context compaction and retry logic, probably. Sandboxing, permissions, spend limits and audit trails, no - those encode what your organization will allow, not what the model can do, and no model release repeals them. Memory is the other durable bet: it is the dimension models still handle badly, and investing there has already beaten a vendor's own harness on that vendor's model.
Our starter kit is built on exactly that filter: instruction files with a ratchet log, a permissions and spend checklist, and a local-model config. Get the kit.
Buyers found this subject before the answers did. In one publisher's Search Console, 48 distinct queries containing "harness" landed on a single page in ninety days - at positions 5 to 20, with zero clicks between them. The category has demand and scattered answers. HarnessKit keeps one maintained comparison, one chooser, and one kit, and cites its sources so you can check us.
Primary source: Dr. Phil Winder, "A Comparison of AI Agent Harnesses in 2026" (Aug 20, 2026; table checked Aug 19). Category timeline: LangChain's harness anatomy (Mar 10, 2026), Sebastian Raschka's components (Apr 4), Addy Osmani's harness engineering (Apr 19), Pydantic AI v2.0.0 harness-first release (Jun 23), DeepSeek Harness developer preview (Aug 13).