ayjnt Get started
Documentation menu
Concept

Harness engineering

Harness engineering is the work of designing the system around a model so an agent can act reliably, safely, and legibly.

The model is not the product

Model quality matters, but most useful agent behavior comes from the environment around the model. A strong harness gives the model the minimum context it needs, tools with clear contracts, durable state, recovery paths, and timely human input.

A practical harness loop

  1. Observe: gather the state, history, and external signals relevant to this turn.
  2. Decide: ask a model or deterministic policy what should happen next.
  3. Act: execute a narrow tool in the appropriate runtime.
  4. Record: persist the result and make it visible.
  5. Escalate: ask a person when intent, authority, or judgment is missing.

Five design principles

01

Constrain context

More context is not automatically better. Supply what the current decision needs.

02

Make tools small

Narrow inputs and outputs are easier to validate, observe, approve, and retry.

03

Separate authority

Reading, writing, and executing on a host should be visibly different capabilities.

04

Design for interruption

Long work needs checkpoints, resumability, cancellation, and human approval.

05

Expose the work

People should see what an agent knows, what it is doing, and where it needs help.

Human interaction is architecture

Human-in-the-loop is not a confirmation dialog added at the end. It changes how work is modeled. A harness needs durable pause points, clear previews of proposed actions, enough context for a person to decide, and a way to resume without repeating completed work.

A useful approval contains a decision

Show the proposed action, its consequences, and the exact scope of the authority being requested.