---
kind: slashhuman-agent-field-manual
version: 1
title: The Harness Is the Agent
slug: the-harness-is-the-agent
summary: Move reliability out of model personality and into durable, testable conditions.
human_keeps: The definition of reliable and the consequences the system may expose.
agent_takes: Harness inspection, failure containment, verification, and durable capture.
---

# Agent field manual: The Harness Is the Agent

Use this when the human describes a model as unusually smart, lazy, forgetful, or
inconsistent, or when a workflow depends on the model remembering how it behaved in
an earlier session.

## Objective

Make useful behavior reproducible across sessions and model changes. Treat your own
output as a proposal inside a system, never as the durable system itself.

## Operating contract

The human defines what reliability means here, what failures matter, and what risk is
acceptable. You locate where continuity currently lives, expose what exists only in
conversation, and move repeatable behavior into deterministic controls wherever
possible.

## Protocol

1. Inventory the harness: source of truth, project instructions, tools, permissions,
   tests, schemas, decision records, deployment gates, and recovery path.
2. Mark every expectation that currently depends on memory, model temperament, or a
   human noticing something manually.
3. For each important expectation, choose the cheapest durable carrier. Prefer, in
   order: executable test or schema, automated check, permission boundary, template,
   written instruction, review checklist.
4. Run the workflow against an adversarial or degraded output. Verify that the
   harness catches the failure without requiring the model to confess it.
5. When you produce an unexpectedly good behavior, propose where to capture it so a
   later agent can reproduce it without reading this conversation.

## Refuse these defaults

- Do not ask the human to trust confidence, style, or a model name.
- Do not solve deterministic validation with another generation step.
- Do not add ceremony whose failure mode is “someone forgot to do it.”
- Do not claim a harness is safe because the happy path passed once.

## Completion evidence

A fresh agent can enter with no conversational history, locate the source of truth,
perform the workflow, and be stopped by the system when it violates a named
condition.

Canonical essay: https://slashhuman.blog/posts/the-harness-is-the-agent/