Domain specific harness
A domain specific harness is that harness that is tuned to one vertical's tools, workflows, failure modes, and edge cases. The idea being if the model is a commodity the value will accrue where the actions are decided and done.
An agent harness is the translation layer between an AI model and the environment it needs to interact with. An agent is the result of pairing the AI model with a harness.
A harness requires at least:
- A system prompt, which explains to the AI model what it is and how it should act.
- Describe a set of tools and capabilities, usually known as skills and MCPs.
- Translation layer to allow it work with different models and their intrinsicallities.
- Agent loops: an AI agent receives a task, uses tools to gather information, checks whether the results are good enough, and repeats steps if needed. Once satisfied, it creates the final output and delivers it to the user.
Fine-tune those four for a specific domain and verticalize (this is also relevant to the digital to physical world interactions).
An important insight about agent harness is how it changes through use. Prompts, memory, skills, and routing rules becoming persistent state that agents can edit.
A domain specific harness can also accumulate experience. An agent might save a useful procedure, change an instruction after a failure, or adjust which tool it uses. Once those changes persist across sessions, they affect every task that follows.
This makes evaluation part of the domain-specific work.
- A customer support agent might learn a faster resolution procedure, but does it still handle exceptions correctly?
- A robotics agent might find a shorter sequence of actions, but does it still respect physical constraints? Harness edits need a promotion gate. The agent proposes a change, an evaluator compares it with the previous version on tasks that did not shape the edit, and the gate decides whether to adopt it.
My bet is that defining those tasks and acceptable tradeoffs becomes a large part of building for a domain. You need to know what success means, which failures matter, and what must keep working as the agent changes.
Each new model is a reason to re-evaluate the harness. Some instructions and workarounds may become unnecessary. Domain-specific evaluations help you decide what to remove and whether the simpler version still does the job.
The bear case for domain a domain specific harness? It's gonna be pretty hard for any harness to outperform a frontier team over time, especially if frontier teams have infinite compute on models that aren't publicly available.