Harness is the software wrapped around a large language model that turns it into an AI agent, covering the tools it can call, the loop that repeats until the task is finished, its memory, and the limits on what it may do.
For example, Claude Code, Anthropic's coding tool, is described in its own documentation as the agentic harness around Claude. The term was popularized after Mitchell Hashimoto, co-founder of the infrastructure company HashiCorp, wrote about engineering the harness on 5 February 2026, though he stated that no accepted term yet existed; Anthropic had already used the word in engineering writing in November 2025.
Which harness an agent runs inside often matters more than which model it uses, since identical model weights produce different results depending on the tools supplied, the context assembled, and the checks applied afterwards. LangChain, a company that builds agent software, reported in March 2026 that it moved its coding agent from outside the top thirty to fifth place on the Terminal Bench 2.0 benchmark by changing only the harness. OpenAI reported in February 2026 that three engineers directing its Codex agent merged roughly 1,500 pull requests over five months without writing any code themselves.
Some critics argue that harness work merely compensates for weaknesses that better models will remove. The writer, swyx, who publishes the AI newsletter Latent Space, set out this position in March 2026, citing Boris Cherny and Cat Wu of Anthropic's Claude Code team, who describe their product as the thinnest possible wrapper over the model. The OpenAI researcher Noam Brown argues that the elaborate scaffolding built for earlier models became unnecessary once reasoning models arrived.
In work automation, the harness sets permission and oversight, because the model proposes an action while the harness decides whether it actually runs.
For example, an agent handling supplier invoices could read the accounting records and prepare a payment, while the harness holds the transfer until a finance manager approves it, which is one of the guardrails enterprise deployments commonly require. Those checks are charged on every run: researchers at ETH Zurich reported in February 2026 that instruction files added to each session raised inference cost by more than twenty percent while reducing task success in five of the eight settings tested.
