Harness Engineering Project Template

A lightweight project skeleton for using Harness Engineering with AI coding agents.

Harness Engineering means surrounding an AI coding agent with:

• Feedforward: specifications, conventions, constraints, and required context.
• Feedback sensors: tests, linters, logs, metrics, review agents, and human review.
• Hooks: the scripts, CI jobs, log access, and workflows that connect the agent to feedback.
• Correction loops: rules for fixing, retrying, escalating, or stopping.

The goal is to make AI-agent work more reliable, reviewable, and governable.

Inspired by Birgitta Böckeler’s article:
https://martinfowler.com/articles/harness-engineering.html

---

Core Idea

Feedforward → Agent Work → Feedback Sensors → Correction Loop

The agent should not work from a prompt alone. It should work inside a harness that defines:

1. what to build,
2. how to build it,
3. which checks to run,
4. how to interpret feedback,
5. when to self-correct,
6. when to escalate to a human.

---

Project Structure

Current repository contents:

.
├── AGENTS.md
├── README.md
├── docs/
│   └── README.md
├── evaluation/
│   └── rubric.md
├── harness/
│   ├── ai-agent-feedforward-template.md
│   ├── feedforward.md
│   ├── harness.yaml
│   ├── correction-policy.md
│   ├── risk-controls.md
│   ├── final-response-contract.md
│   ├── sensors/
│   └── hooks/
├── scripts/
│   ├── agent-preflight.sh
│   ├── run-computational-sensors.sh
│   └── agent-final-check.sh
├── src/
│   └── README.md
├── tasks/
│   ├── task-template.md
│   └── examples/
│       └── example-bugfix-task.md
└── tests/
	└── README.md

---

Main Files

File	Purpose
AGENTS.md	Main entrypoint for AI agents
harness/feedforward.md	Defines what to build, how to build it, and how to verify it
harness/ai-agent-feedforward-template.md	Optional AI-agent feedforward prompt template
harness/harness.yaml	Machine-readable harness configuration
harness/correction-policy.md	Defines fix, retry, escalation, and stop rules
harness/risk-controls.md	Defines autonomy limits by risk level
harness/final-response-contract.md	Defines what the agent must report when done
tasks/task-template.md	Template for task-specific feedforward
scripts/run-computational-sensors.sh	Placeholder command entrypoint for tests, lint, typecheck, and build
evaluation/rubric.md	Rubric for evaluating agent output and harness quality

---

How to Use

1. Copy this template into a new repository.
2. Fill out harness/feedforward.md with project-specific product and engineering guidance.
3. Configure harness/harness.yaml with real commands for your stack.
4. Create a task file from tasks/task-template.md.
5. Customize scripts/run-computational-sensors.sh with your stack commands.
6. Point the AI agent to AGENTS.md.
7. Require the agent to run the relevant sensors and follow the final response contract.

---

Agent Workflow

Read AGENTS.md
→ Read harness feedforward
→ Read task feedforward
→ Inspect required context
→ Make the smallest safe change
→ Run computational sensors
→ Perform inferential checks
→ Apply correction policy
→ Report results

---

Computational and Inferential Sensors

Computational sensors are executable or measurable checks:

• unit tests,
• integration tests,
• type checks,
• linters,
• build checks,
• static analysis,
• logs,
• metrics.

Inferential sensors require judgment:

• requirement alignment review,
• architecture fit review,
• test quality review,
• risk assessment,
• human review,
• AI code review.

Both are needed. Computational sensors catch objective failures. Inferential sensors judge whether the work makes sense in context.

---

Status

This is a starter template. Adapt it to your language, framework, CI system, observability tools, and team conventions. And LLM Resources such as context window