Agent harness: the layer that turns a model into a dependable agent
The engineering layer wrapped around a language model — retrieval, tools, skills, memory and guardrails — that turns a general model into a specific, dependable agent. The model supplies fluency; the harness supplies grounding, boundaries and the ability to act, and it stays the same when the model underneath is swapped.
A language model, on its own, is fluent and unbounded. It will answer anything, from its training data, in a plausible voice — including the things it should refuse, the facts it no longer has, and the price you changed last week. A harness is everything you wrap around that model to make it useful for a real job: the retrieval that grounds it in your material, the tools and skills that let it do things, the memory that carries a conversation, and the boundaries that stop it saying what it must not.
The word comes from how the term is used by the people who build these systems. The model is the engine; the harness is what turns raw horsepower into something that pulls a specific load, in a specific direction, without bolting.
What a harness is made of
The pieces have their own names on this site, and a harness is how they fit together:
- Grounding and retrieval — answers are drawn from your documents, not the model's memory or the open web. This is what separates an assistant that cites your policy from one that sounds like it read a policy once.
- A soul — the standing instructions and the concrete refusal line: never quote a rate, never predict an outcome, never assess a case. Written per agent, checked by a human.
- Skills and tools — so the agent can collect a document, book a slot or file a record, not only talk about doing it.
- Memory and continuity across a customer's visits, so they are not starting from zero each time.
None of these are the model. They are the layer that makes the model behave.
Why the harness, not the model, is the thing you keep
Models are becoming a commodity — open weights, a new frontier release every few months, an endpoint you can swap for your own. When the intelligence is interchangeable, the durable, defensible work moves to the layer around it. A harness that grounds well, refuses reliably and can be handed to a non-engineer to configure for their trade is worth more the stronger models get, because a stronger model needs more, not less, holding to a line.
That is the whole bet of a business agent: the value is not a smarter model, it is a model kept honest, on-topic and inside your rules.