Does your AI coding agent regularly "confidently deliver a pile of code that doesn't run"? The problem most likely isn't the model itself — it's whether you've fitted it with a Harness.

Anthropic and OpenAI, almost simultaneously in 2026, validated the same conclusion with experiments: when AI coding agents keep failing, the problem isn't the model — it's the Harness infrastructure outside the model. The same Opus 4.5 model, run bare, burned $9 and failed every time; fitted with a Harness, it spent $200 and hit a 100% success rate.

What Is a Harness

A Harness isn't a tool, and it isn't a prompting trick — it's the complete engineering infrastructure built around an AI coding agent, made up of five subsystems:

Two Controlled Experiments

The Anthropic Experiment

The same Opus 4.5 model, the same coding problem:

• Run bare: $9 spent, 0% success — messy code style, destructive commands, no tests run
• With a Harness: $200 spent, 100% success — the extra $191 all went into the verification loop

The OpenAI Experiment

The Codex team validated this on real repositories with millions of lines. The experiment changed exactly one thing -- an AGENTS.md file, under 100 lines of markdown, added at the repository root.

Build Your Harness in Five Steps

The five steps below can all be done in a text editor, adding up to no more than 200 lines of configuration.

Step 1: Create AGENTS.md (or CLAUDE.md)

Create a markdown file in the repository root. The OpenAI camp calls it AGENTS.md; the Anthropic camp calls it CLAUDE.md. Codex, Claude Code, and Cursor read it automatically at startup and inject it into the system prompt.

Write at least three blocks of content:

# Project
This is an e-commerce admin backend built with Next.js + Prisma

# Forbidden
- Never run git push --force
- Never delete the migrations directory
- Never use npm (use pnpm)

# Done means
- pnpm typecheck passes
- pnpm test all green
- pnpm lint zero errors

In under 15 lines, project conventions go from something you repeat over and over to something auto-injected at startup.

Step 2: Configure Permissions

Restrict which commands the agent can invoke.

Claude Code uses .claude/settings.json; Codex uses ~/.codex/config.toml.

{
  "permissions": {
    "allow": ["pnpm install", "pnpm test", "pnpm typecheck"],
    "deny": ["rm -rf", "git push --force", "DROP TABLE"]
  }
}

Allowed commands run directly, denied commands are refused outright, and gray areas pop a confirmation.

Step 3: Write setup.sh to Lock the Environment

Lock down dependency versions and runtime configuration. If you already have a Dockerfile / devcontainer.json you can skip this; otherwise write a setup.sh.

The key line:

pnpm install --frozen-lockfile

--frozen-lockfile ensures the agent can't upgrade any dependency on its own.

Step 4: Create PROGRESS.md

touch PROGRESS.md

Four sections: completed, in progress, to do, known issues. Commit it to git and maintain it as part of the project itself.

Codify the convention in AGENTS.md:

## Rules
- First thing in a new session: read PROGRESS.md
- On task completion or checkpoint change: write back to PROGRESS.md immediately
- On conflict, code wins -- the repo is the single source of truth

Step 5: Codify the Definition of Done (Most Critical)

State the verification commands at the end of AGENTS.md:

## Done Definition
Task is NOT done until ALL of these pass:
- pnpm typecheck (exit code 0)
- pnpm test (exit code 0)
- pnpm lint (exit code 0)
- pnpm build (exit code 0)

If an exit code isn't 0, the task isn't done. If your project doesn't have these commands yet, set them up today.

Core lesson: get the first four steps right but skip the fifth, and it's still all wasted. Without a feedback loop, a Harness might as well not be installed.

Three Deadly Failure Modes

The Anthropic and OpenAI experiments point to the three most common agent failures:

1. Declaring Victory Too Early

The agent finishes 500 lines of features and outputs "done." Merge the code -- CI goes red, type check reports 12 errors, and not a single unit test has run.

Fix: the feedback subsystem. Hand the verdict to exit codes -- exit code != 0, task != done.

2. Context Anxiety

A long task hits 70%, and context Tokens are nearly full. The agent starts rushing -- skipping tests, deleting edge-case handling, wrapping things up with stubs.

Fix: the state subsystem + proactive restarts. When context Token usage passes 70%, deliberately stop, finish writing the checkpoint, and start a new session.

3. Cross-Session Amnesia

The first session writes the user module; the second session writes getUserById all over again, and the interface signatures conflict.

Fix: PROGRESS.md maintains the completed-features list + AGENTS.md states the read-it-first convention.

The Core Conclusion

Model capability sets the ceiling; the Harness decides how much of that ceiling you actually reach.

Without a Harness, Opus 4.5 produces code that can't even compile; with one, even a tier-smaller model delivers reliably. Rather than waiting for the next stronger model, install the Harness first.

Resources:

• Learn Harness Engineering
• Everything Claude Code (ECC)