The model thinks; the harness does everything else. This page explains what a harness is, the kinds in use today, how much of Y Combinator's 2026 output is harness-shaped, and what we expect to come next.
Researched October 2026. Claims about other people's work cite a source (S1 to S16, listed at the end). The YC figures are our own count from YC's public directory; the method and its limits are written out. The last section is a prediction and is labelled as one.
A large language model on its own is a function: text in, text out. It has no memory between calls, can't run a command, and doesn't know when it's finished. An agent harness (also called scaffolding) is the software around the model that supplies all of that: the loop that calls the model again and again, the tools it can use, the memory and state that persist between calls, the sandbox it works in, and the checks that tell it whether it succeeded. The shorthand that has stuck is agent = model + harness. S1 S2
A typical harness has these parts: S1 S3
The word itself is older than LLM agents. A test harness runs code under controlled conditions and an evaluation harness runs a model against a benchmark; both meanings are still in use (section 2). S1
Researchers studying coding agents now argue that you can't judge a model without its harness. One 2026 study held the model fixed and ran 35 consecutive releases of the Qwen Code CLI on the same 50 SWE-bench Verified tasks, to show that quality changes from release to release of the harness alone; the five open-source harnesses it surveyed shipped more than two releases a day. S4
In February 2026 the practice got a name. Mitchell Hashimoto (co-founder of HashiCorp) wrote that he'd "grown to
calling this harness engineering": every time an agent makes a mistake, fix the environment so it can't make it again,
by updating AGENTS.md or by writing a tool that gives it fast, reliable feedback. S5
Six days later OpenAI published an account of a team that built an internal product of about a million lines with no
hand-written code, describing the engineers' job as designing the environment while "agents execute". S6
In April, Birgitta Böckeler of Thoughtworks gave the clearest model so far: a harness is made of guides
that steer the agent before it acts and sensors that check its work afterwards, and each can be
computational (tests, linters, type checks) or inferential (an LLM reviewing the output). S7
Our own setup is an example. The AGENTS.md that steers our coding agents is a guide; the tests and the observability stack it requires are sensors.
| Kind | What it does | Examples |
|---|---|---|
| Coding agent harnesses | Wrap a model in a loop with shell, file and git tools so it can change a codebase and test its own work. | Claude Code (Anthropic), Codex CLI (OpenAI), Gemini CLI (Google), Qwen Code, OpenCode, OpenHands, Cursor, Antigravity (Google DeepMind) S1 S4 |
| General agent harnesses | The same loop, for work outside code: research, operations, personal assistants. | OpenClaw (Peter Steinberger), LangChain Deep Agents, Claude Agent SDK, OpenAI Agents SDK S8 S9 |
| Evaluation harnesses | Run a model or agent against a benchmark under fixed conditions and score it. The older meaning of the word. | EleutherAI lm-evaluation-harness (the backend of Hugging Face's Open LLM Leaderboard), UK AI Security Institute's Inspect, SWE-bench's evaluation harness S10 S11 |
| Frameworks and orchestration | Libraries for building your own harness, or for wiring several agents and deterministic steps into one system. | LangGraph, CrewAI, Microsoft AutoGen, Mastra S9 |
| Outer harnesses | What a team adds around a vendor's agent to fit its own project: instructions, skills, scripts, checks. | AGENTS.md / CLAUDE.md files, Skills, CI checks, sandboxes; Böckeler's "harness for coding agent users" S5 S7 |
| Wrapper products | A harness for one job or industry, sold as the product: the model is rented, the loop, tools, integrations and data are the business. | Most AI startups in section 3, from AI accounting firms to agents for permitting or penetration testing |
Short answer: about half of YC's 2026 companies sell a harness as their product, and roughly six in ten are either a harness or a part for one. Nobody publishes an official count, so we made our own.
PitchBook counted AI agent companies among YC's Demo Day presenters: 58 of 163 in Winter 2025 (about 36%) and 67 of 144 in Spring 2025 (46%). S12 Nobody we found has counted "wrappers" or "harnesses" directly; claims online that "90% of YC AI startups" build on one API have no source behind them, so we don't repeat them.
We pulled every company in each batch from YC's public directory (via the open yc-oss mirror, updated
9 October 2026) and searched each one-liner and description. S13
| Batch | Companies | Mentions AI | Mentions "agent" | "Agent" in the one-liner |
|---|---|---|---|---|
| Winter 2025 | 165 | 80% | 31% | 13% |
| Spring 2025 | 144 | 85% | 40% | 22% |
| Summer 2025 | 166 | 90% | 43% | 20% |
| Fall 2025 | 145 | 85% | 43% | 21% |
| Winter 2026 | 198 | 79% | 36% | 19% |
| Spring 2026 | 193 | 83% | 49% | 26% |
| Summer 2026 | 231 | 83% | 34% | 16% |
| Fall 2026 (in progress) | 123 | 80% | 36% | 25% |
Keyword matches over-count a little (an "AI-native" law firm counts as AI) and under-count a lot (many agent products never use the word). They show the trend, not the answer. 22 companies across these batches describe themselves with the word "harness", among them Moda ("harness engineering platform") and Hilstart ("physical agent harnesses").
To get closer to the real question, we drew a random sample of 150 of the 622 companies in Winter, Spring and Summer 2026 (fixed seed, so it can be re-run) and read each description, sorting it into one of four groups:
| Group | What counts | Companies | Share |
|---|---|---|---|
| Harness products | An agent, copilot or "AI-native" service for a job or industry, built on rented models: accounting, legal, DevOps, sales, permitting | 71 | 47% |
| Harness infrastructure | Parts for other people's harnesses: context compression, agent permissions, sandboxes, phone numbers and payments for agents, agent memory | 18 | 12% |
| Models, compute and data | Trains its own models, sells inference, or sells training data to labs | 18 | 12% |
| Everything else | Robotics, hardware, space, biotech, fintech and consumer products where AI isn't the product | 43 | 29% |
With a sample this size, the true share of harness products in those three batches is likely between 40% and 55%, and harness products plus harness infrastructure between 51% and 67% (95% intervals). Spring 2026 was the most harness-heavy batch in the sample (31 of 52); Summer 2026 the least (15 of 41), with more hardware, robotics and space.
Limits, stated plainly. One reader (us) sorted the sample from YC's own short descriptions, which are marketing copy. Some calls are judgement: an "AI-native" law or accounting firm counts as a harness product because its leverage is agents on rented models, even though it also sells human work. "Harness" here describes the architecture, not quality: many of these firms own hard-won integrations, data and customers that a thin wrapper wouldn't. A second, independent reader would make this stronger.
The economics point the same way. In May 2026 OpenAI offered every company in YC's spring class $2 million in tokens in exchange for an uncapped SAFE, which TechCrunch read as a bid to have the batch build on its models. S14 When the model is subsidised and swappable, what a startup can own is the harness around it: the workflow, the integrations, the data, the checks and the customer. That is also why we keep at least two model providers behind an adapter (own your stack).
All from YC's 2026 directory, in the companies' own words, to show the range. S13
The industry has named a new layer every few months: prompt engineering, then context engineering (what goes into each call), then harness engineering (the environment around one agent), then loop engineering (agents that run themselves on a schedule until a goal is met). S8 The newest name is graph engineering.
Graph engineering describes a whole system as a graph. Nodes do the work, and a node can be plain code, one model call, a tool call, a human approval, or a complete agent with its own harness. Edges decide what runs next, either fixed or chosen at run time. The graphs usually have cycles, for retries, validation and waiting on a person. S9 The term took off on X on 18 and 19 July 2026 after a post by Peter Steinberger, the creator of OpenClaw, about loops versus graphs, and had several competing definitions within two days. S8 We cover it in depth on a companion page, graph engineering. In August a 35-author survey framed it as the step from "individual intelligence" to "system intelligence": explicit, changing graphs of tasks, agents and state, for work too big for one agent to organise. S15
We think graph engineering is real but is not the next harness. The harness was a new product; the graph is mostly a new name for wiring harnesses together. Here is what we expect over the next 12 to 24 months, with how sure we are:
What this means for us: keep our harness thin and our adapters many (Claude Code, Codex, Kimi and a local model behind one interface), put our effort into guides and sensors, and treat graphs as a tool for known workflows, not as the architecture.
Provided as is, for information only, with no warranty and no liability. Not legal or professional advice. Security and radio material is for authorized, lawful use only. Full disclaimer.
Visits are logged on our own server: IP address, browser and device details, and pages viewed. We use them to understand traffic and spot attacks, and never share or sell them.