Harness engineering

Harness engineering is the design of the tools and workflows that let a language model work on multi-step tasks. A harness manages model and tool calls, context, state, and access to the working environment. It also provides task-specific instructions, checks, and records so people can review the work.

The parts of a harness

  • Execution

    Model and tool calls, files, the command line, network access, permissions, and error handling.

  • Skills

    Instructions for a particular task: what to do, in what order, and how to check the result.

  • Checks and gates

    Checks provide evidence for a decision. A gate stops the workflow when a required condition is not met.

  • Context and state

    Project conventions, boundaries, relevant lessons, and the current state of work across sessions.

  • Process

    A repeatable path from requirements to a decision, a plan, implementation, and verification.

  • Evidence and records

    Check output, measurements, and a record of actions that a person can inspect before accepting the work.

  • Editor adapters

    In this toolset, adapters bring a shared set of skills into different coding environments.

Make a failed check matter

An instruction can tell an agent to check its work. A gate makes meeting a specific condition part of the workflow.

Choose a check that fits the task. For a code change, it might be a test; for a research note, it might be a requirement to cite sources. Save the actual output so another person can review it.

ILLUSTRATIVE RESEARCH CHECK

Every factual claim needs a source.

A draft contains two claims without source links.

Decision: stop and resolve the missing sources.

The check result stays with the draft. The reviewer can see what failed and what changed.

Common questions

How is this different from a prompt?

A prompt gives instructions to the model. A harness includes those instructions alongside execution, tools, context, state, checks, and records. The full system determines how the agent works and how people review its output and decide whether to accept it.

Is harness engineering the same as agentic engineering?

Here, harness engineering is one part of the broader practice of agentic engineering: defining tasks, dividing work between people and agents, and accepting results. Industry usage varies; there is no single settled boundary between the terms.

Where should I begin?

Choose one recurring task and write down what would make its result acceptable. Add one relevant skill, one meaningful check, and a place to keep the output. Expand the harness when the work calls for it.

Put the idea to work

Start with one task and decide how you will check the result.

Explore the tools