What people mean by a harness

A model on its own takes text in and gives text out. What makes it do a job is everything around it: what it is told first, the tools it can call, the checks that decide whether the work passed, what it remembers, what is recorded, and who gets asked. Harness is the word that has caught on for that system.

I can source four uses. Mitchell Hashimoto wrote in February 2026 that whenever an agent makes a mistake he takes the time to engineer a fix so it never makes that mistake again, and that he had started calling this "harness engineering" [Hashimoto 2026]. He said he was not sure the industry had a name for it, and that he would use another if one existed [Hashimoto 2026]. He describes two kinds of fix: better instructions in an AGENTS.md file, and programmed tools such as scripts [Hashimoto 2026].

Anthropic’s engineering team uses the word for the system around an agent. A post of November 2025 calls its Claude Agent SDK "a powerful, general-purpose agent harness", and describes one built so that each fresh session can pick up the last one’s work: a feature list, a progress file, git commits and a start-up script [Anthropic 2025 (vendor)]. A post of March 2026 on a planner, generator and evaluator design says "every component in a harness encodes an assumption" about what the model cannot do alone, and that such assumptions can go stale as models improve [Anthropic 2026 (vendor)]. Neither gives a definition, and both are a company describing its own approach.

OpenAI’s post of February 2026 has harness in its title and uses it elsewhere only for the "evaluation harnesses" its agents write, and does not define it [OpenAI 2026 (vendor)]. Its subject is a team that wrote no code by hand, whose job became designing environments, specifying intent and building feedback loops [OpenAI 2026 (vendor)]. When the agent struggled, the team asked what was missing (tools, guardrails or documentation) and fed it back into the repository [OpenAI 2026 (vendor)]. It treats AGENTS.md as a short map to a docs directory, not an encyclopedia [OpenAI 2026 (vendor)]. It also hedges: the approach depends on that repository’s structure and tooling and should not be assumed to generalise [OpenAI 2026 (vendor)].

One contrast with this site, stated plainly. The OpenAI post describes a repository with minimal blocking merge gates, where a person may review a pull request but is not required to, and it says that would be irresponsible in a low-throughput environment [OpenAI 2026 (vendor)]. The pages linked below argue for gates in code and a person’s approval before anything that cannot be undone. I report the difference and do not settle it here.

So the honest summary is short. The word is recent, the sources I could read use it for the system around the model, and none of them defines it. I do not credit it to anyone.

The parts below are my grouping, not a standard one, and the sources above name some of them and not others. Each links to the page that covers it.

• The loop and its gates

  One loop, three modelsWrite down what done looks like as checks, then build until code says they pass, with a person before anything ships.

• The graph around the model

  Designing an agent graphModels, tools, routers, gates, memory and people as a graph, in six patterns, the last of them about failure and budgets.

• Memory

  The vault and the graphA folder of notes as the memory of an agent graph, with and without it.

• Records of a run

  Traces, checkpoints and receiptsWhat happened, where to resume, and proof that a check passed on exactly this content.

• Who owns it

  The next platform teamWho provides the gateway, the tool access and the permissions the graphs run on.

• Gating an action with a probability

  Gating an agent’s actionChecking an agent’s proposed action before it runs, and why a refusal is not the same as asking a person.

A harness is not fixed. The same Anthropic post says to re-examine a harness, stripping away pieces that are no longer needed and adding new ones for what the model can now do; it reports dropping the sprint construct and context resets from its own design when a newer model arrived [Anthropic 2026 (vendor)]. Whether a given piece earns its place is a question for your own tasks and your own measurements, not for a slogan.

• The word is young and used differently by different people. Treat each use above as that author’s, not as a definition.
• Three of the four sources are a company (Anthropic twice, OpenAI once) describing its own approach. The fourth is one engineer’s account of his own habits. None is a measurement, and the OpenAI post’s own figures are not repeated here.
• The map of parts is my own grouping.

• Hashimoto 2026 · Mitchell Hashimoto, "My AI Adoption Journey", mitchellh.com (2026-02-05; one practitioner’s account), the step "Engineer the Harness". Read 2026-11-01
• Anthropic 2025 (vendor) · J. Young, "Effective harnesses for long-running agents", Anthropic Engineering (2025-11-26; a vendor describing its own approach). Read 2026-11-01
• OpenAI 2026 (vendor) · R. Lopopolo, "Harness engineering: leveraging Codex in an agent-first world", OpenAI Engineering (2026-02-11; a vendor describing its own team’s experiment). Read 2026-11-01
• Anthropic 2026 (vendor) · P. Rajasekaran, "Harness design for long-running application development", Anthropic Engineering (2026-03-24; a vendor describing its own approach). Read 2026-11-01