Before the Harness Had a Name

In October 2022, researchers from Princeton University and Google Research posted a paper called ReAct. Its idea fits in a sentence: let a language model alternate between reasoning in words and taking an action, then feed the result of that action back in before it reasons again.

Thought, action, observation, repeat. On question-answering tasks the model could look things up through a tiny search tool instead of relying on what it half-remembered, and on two interactive decision-making benchmarks the paper reported gains of 34 and 10 percentage points in success rate over the methods it compared against.

Look closely and you can see a harness in miniature. Someone had to write the code that spotted an action in the model's output, ran the search, trimmed the result and pasted it back into the prompt. Someone had to decide when the loop should stop. The paper does not call that code a harness. But every system in the rooms ahead descends from that small program running beside the model: the part that turns text into consequences, and consequences back into text.

Two years later, in December 2024, Anthropic published a short guide called Building Effective AI Agents, and gave the field a distinction it still uses.

In a workflow, the path is written in code: the model is called at fixed steps, and the program decides what happens next. In an agent, the model directs its own process, choosing which tool to use and when it is finished, with feedback from the environment at every step. The guide's advice was deliberately unglamorous. Start with the simplest thing that works, add autonomy only when the task needs it, and spend real effort on the interface between the model and its tools, documenting them as carefully as you would for a new colleague. Customer support appears in it as a natural fit for agents: the conversation flows, but the actions can be checked.

The same month, Hugging Face's announcement of its smolagents library described agency as a spectrum rather than a switch: from a model whose output changes nothing about what the program does, through a router and a tool caller, to a loop the model controls and, at the far end, one agent starting another. Put the two objects side by side and a question appears that every harness now has to answer, one task at a time: who draws the path, the code or the model?

By 2025 the hard problems had moved from the model to what the model was shown.

In June, two essays published days apart seemed to disagree. Cognition, which builds a coding agent, argued against splitting work across parallel sub-agents: each helper makes quiet decisions the others cannot see, and the pieces stop fitting together. Share the full trace, it said, or keep one thread. Anthropic described a research system that did the opposite, a lead agent sending sub-agents to search in parallel, and reported that it beat a single agent by 90.2% on its internal research evaluation while using roughly fifteen times the tokens of a chat (vendor-reported). The dialogue at the end of this room reconciles them.

In July, the team behind Manus published what it had learned about context in production: treat the cache hit rate as a key metric, keep the opening of the prompt stable and the history append-only so cached work is reused, and leave failed attempts in view so the model does not repeat them. The same month, Chroma's Context Rot report tested 18 models and found every one degraded as its input grew longer, even on simple tasks (vendor research). In September, Anthropic named the practice context engineering: finding the smallest set of high-signal tokens for each step, using compaction, notes kept outside the window, and retrieval just in time.

The lesson of the room is that a context window is not storage. It is a budget, and the harness keeps the accounts.

As models got better at longer tasks, a new problem surfaced: the job outlived the context window.

METR, a research group that measures what AI systems can do, reported in March 2025 that the length of tasks models could complete, measured in the time they take a skilled person, had been doubling roughly every seven months. Its January 2026 update, Time Horizon 1.1, reported that the pace had quickened. A job that lasts hours will not fit in one conversation, so something outside the model has to carry it.

LangChain's Deep Agents post, in July 2025, listed what the most capable agents of the day had in common: a planning tool, sub-agents, a file system to write to, and a long, detailed prompt. In November, Anthropic described a harness for agents that work across many sessions. An initializer session sets up the project and writes a feature list and a progress file; each later session reads the notes, picks one feature, tests it, updates the notes and commits. The essay compares each session to an engineer arriving for a shift with no memory of the last one. The fix was not a bigger memory. It was a better handover.

A follow-up in March 2026 put the general lesson plainly: every part of a harness encodes an assumption about what the model cannot yet do, and those assumptions go stale as models improve. This is the room where memory stops meaning what is in the prompt, and starts meaning what the harness writes down, where, and for whom.

The word was already circulating. In January 2026 Philipp Schmid described the model as a computer's processor and the harness as its operating system. February made it a discipline.

On February 5, Mitchell Hashimoto wrote about his own adoption of AI tools and gave one step a name, engineering the harness: whenever an agent makes a mistake, change its environment so that mistake cannot happen again. Six days later OpenAI published an account of software built with no hand-written code, roughly a million lines across about 1,500 pull requests written by agents (vendor-reported), steered through a short instructions file that works as a map to deeper documentation. On February 17, Birgitta Böckeler sketched harness engineering on Martin Fowler's site as context engineering plus architectural constraints plus regular clean-up. In April she sorted the controls into guides, which steer an agent before it acts, and sensors, which check its work afterward, each either computational or inferential.

By March, LangChain was writing the equation outright: agent equals model plus harness. Wikipedia's entry, as edited in September 2026, records that the term's origin is contested, and distinguishes an inner harness, the loop a product builds around its model, from an outer harness, the rules, checks and knowledge a team builds around that. Most businesses will never touch the inner one. The outer one is theirs.

When a practice gets a name, institutions follow.

In December 2025 the Linux Foundation formed the Agentic AI Foundation to give the Model Context Protocol, the goose agent framework and the AGENTS.md convention a neutral home. The same month OWASP published its Top 10 for Agentic Applications, a list of risks that opens with goal hijacking and includes memory and context poisoning: the dangers of an agent that reads untrusted text. In February 2026, NIST announced an AI Agent Standards Initiative centred on something harnesses had been improvising, giving agents their own identity and authorization instead of a borrowed, generic service account. And on July 28, 2026, the final Model Context Protocol specification shipped with a stateless core, moving long-running Tasks into an extension.

None of these documents tells a business what its assistant should know or say. What they do is fix the vocabulary for the parts around the model: who the agent is, what it may touch, how its tools are described, and what to fear. The harness had become something you could write a standard about.