The Agent Harness: Building reliable AI workflows

Everyone who has built something real with language models knows the moment: the demo works, the second run doesn’t. The model is not the unreliable part; the system around it is. That system is what I call the agent harness, and it decides whether an AI workflow is a demo or a product.

What the harness actually is

An agent is only as good as the environment it operates in. I think about that environment in four layers:

• Context. What the agent knows: instructions, domain knowledge, conversation state, constraints. Context that is missing gets guessed, and guesses compound.
• Orchestration. How work is structured: which steps run in which order, what happens in parallel, what retries and what fails fast.
• Tools. What the agent can actually do: narrow, well-described, predictable interfaces beat broad, clever ones.
• Verification. How we know the output is right: assertions, schema checks, evaluations, human review where it matters.

Reliability comes from the edges

The interesting insight from agent engineering is that reliability is mostly an engineering problem, not a prompting problem. The edges of the system (the tool interfaces and the verification layer) carry most of the weight.

// Narrow, well-described tools beat broad ones.
const tool = {
  name: 'get_invoice_status',
  description: 'Returns the processing status of one invoice by ID.',
  parameters: z.object({
    invoice_id: z.string().uuid(),
  }),
};

A tool like this is boring, testable and hard to misuse. Give the same agent a free-form “do admin things” endpoint and watch accuracy collapse.

Feedback loops close the gap

No harness is right on the first attempt. What separates systems that improve from systems that degrade is a feedback loop: outputs get evaluated, failures get classified, and the classification flows back into context, tools or verification. Plan, execute, evaluate, improve: a cycle, not a one-way street.

The mental model

I use this diagram to keep the layers straight:

• Human intent defines goals, instructions and constraints.
• The harness structures what the agent knows, can do and must verify.
• Execution produces results and feedback that improve the next run.

None of this diminishes the model. It is exactly what makes a capable model dependable, the same way a good testing harness makes a capable engineer’s code trustworthy.