Inside the harness: the machinery that turns a model into a worker

In plain terms

Imagine hiring a brilliant new employee and giving them no laptop, no badge, no logins, and no onboarding. Nothing gets produced — not because they aren’t smart, but because they have no way to act. The harness is the laptop, the badge, and the onboarding for an AI model.

And once the employee can act, a second thing matters: how they work. Do they take one sensible step, check the result, and adjust? Or do they disappear for a day and return with a mountain of untested work? Harnesses face exactly this choice, and the good ones force the first behavior.

The anatomy

The core of every harness is the loop. Plan: the model looks at the goal and the current state, and decides the next action. Act: the harness executes the chosen tool — edit the file, run the tests, query the database. Observe: the result (test output, error message, file contents) goes back into the model’s context. Repeat — until the exit condition is met.

Three design choices separate serious harnesses from fragile ones. First, the exit test: what counts as done? “The model says it’s done” is the weakest possible answer; a real harness checks work against something outside the model — tests, rules, an explicit definition of done. Second, tool design: a few well-described tools beat dozens of vague ones, because every tool is another way to act and another way to fail. Third, bounds: iteration limits, time limits, and scope limits, so a confused agent stops instead of looping forever or wandering into systems it shouldn’t touch. Add observability — a full log of every step — and you can debug the agent like any other system. Without it, you’re guessing.

What good looks like

Ask of any agent product: what is the exit condition, and who defines it? What happens when a step fails — retry, replan, or escalate to a human? Is there a complete action log? What are the hard limits? If the answers are vague, the harness is a demo.

Where Cortex stands

In Cortex, the loop runs inside a Mission — a bounded unit of work with a goal, a plan, and a definition of done set before the loop starts. The exit condition is never the model’s opinion; it’s the validation contract, checked with evidence. The standards those checks draw from live in the Kinetic Library.

Questions this paper answers

What are the components of an agent harness? A planning loop, a toolset, state/context management, guardrails, and — in production systems — logging, limits, and verification. The loop decides actions; the rest keeps actions useful and safe.

Why do AI agents get stuck or go in circles? Usually a loop problem: no external exit test, unbounded iterations, or results that never inform the next step. Good harnesses verify work externally, cap iterations, and force small, checked steps.

What tools does a coding agent actually use? Typically: read/edit files, run commands and tests, search the codebase and the web, and query project systems. The craft is giving few, sharply described tools — not many vague ones.