The formula everyone in this industry finally agreed on in 2026 fits on an index card: agent = model + harness. Mitchell Hashimoto put it that way, and the phrase stuck because it names the thing that was always true — the model is a stateless token predictor, and everything that makes it an agent lives in the layer wrapped around it. There's now a whole glossary and an emerging discipline — harness engineering — devoted to that layer.
So what is it?
Everything except the model
The harness is what turns "text in, text out" into work you'd trust: which tools exist and how they're dispatched; what enters the context window and what gets compacted out; what persists between calls, between runs, between days; when the system stops and asks a human; what gets logged, streamed, and audited. Strip the model out of any serious agent product and the harness is what's left — which is to say, almost all of it.
That's also why the model race matters less than it looks from the outside. Frontier models are converging, and they're swappable — a dropdown, an API key. The harness is where agents actually differ, because it's where the unglamorous engineering lives.
What ours does
Auteryn is a harness company. Every item on this list describes the product as it runs today:
- A real machine per agent. Not a tool sandbox conjured per call — a persistent computer with a terminal, a Chromium browser, and a filesystem, which you can open and type into while the agent works.
- Checkpoint after every step. State is persisted before the next action starts. A run that dies comes back with its plan, tool history, and notes intact.
- Context under management. Compaction when the window fills, a working-memory scratchpad, and per-agent tool surfaces so the window holds the task, not the catalog — the subject of its own post.
- A human in the loop, structurally. Six ways a run stops and asks — free text, choice, dropdown, file upload, approval, sign-off on a risk-tagged plan — plus a per-agent autonomy dial.
- Permissions at the API-method level. Allowed to post, denied to delete. Credentials encrypted, tokens scoped and short-lived, every action in the audit log.
- Observability as a feature, not a log file. The live plan, the model's reasoning, and every tool call stream into a timeline you can inspect during the run — and step into.
And the tell, the thing that only makes sense if the harness is the product: the model is your choice, per agent — Gemini, Claude, GPT, or open-source. We don't need to win your model allegiance. Whichever one you pick, it shows up to a machine that's already running, with its context curated, its permissions scoped, and a seat beside it for you.
Why this is the durable layer
Models will keep leapfrogging each other; that's their vendors' race, and everyone downstream benefits. But checkpointing, context discipline, approval gates, audit trails, a machine that persists — that layer doesn't get obsoleted by the next model release. It gets more valuable, because a stronger model on a serious harness is a better agent, while a stronger model on a flimsy one is just a faster way to wander.
The model gets the headlines. The harness does the work.
Auteryn is in private beta — the free plan is a real agent on a real machine, 1,000 credits a month.