A hands-on framework for producing high-quality, maintainable software with AI agents.
"Software engineering is programming integrated over time." — Titus Winters, Tom Manshreck & Hyrum Wright, Software Engineering at Google (2020)
When AI agents write the code, the engineering discipline does not disappear — it shifts into the harness: the system of instructions, guardrails, feedback loops, and documentation that shapes agent behaviour.
This repository is the working toolkit that accompanies the paper Agentic Software Engineering. You can read the paper as a practitioner guide, or hand this toolkit to your AI agent and say: "Apply this framework to my project."
| Path | What it is |
|---|---|
templates/CLAUDE.md |
Annotated agent instruction file template |
templates/AGENTS.md |
Equivalent template for non-Claude agents |
templates/pyproject.toml |
Python quality toolchain (ruff, mypy, bandit, pytest) |
templates/.pre-commit-config.yaml |
Pre-commit hooks |
templates/.github/workflows/ci.yml |
3-job CI: lint + test + security |
templates/.github/workflows/release.yml |
Label-driven semantic versioning + auto-release |
templates/.claude/agents/senior-reviewer.md |
Mandatory pre-merge review agent (adopt and adapt) |
templates/.claude/skills/change-control/SKILL.md |
Change-control runbook: branch, gate, draft-PR, manual-test, release discipline |
templates/.github/pull_request_template.md |
PR template with the sovereign manual-testing checklist |
bootstrap/bootstrap-prompt.md |
Give this to your agent to bootstrap SE quality in any project |
bootstrap/checklist.md |
Post-bootstrap verification checklist |
The paper itself is published separately (see The paper below).
Vibe coding ships fast. It doesn't ship sustainably.
This framework adds six engineering layers around your AI agent:
- Context Engineering — A living
CLAUDE.mdthat tells the agent how to work in your project - Quality Gates — ruff + mypy + bandit + pytest, configured in
pyproject.toml, enforced locally and in CI - Documentation as Code — arc42 architecture docs the agent can read, update, and navigate
- CI/CD Pipeline — Three parallel jobs (lint, test, security) that catch regressions before merge
- Human-Agent Workflow — Planning mode first, feature branches, human approval gate, context clearing
- Entropy Management — Golden principles, error documentation, automated cleanup
The result: an AI agent that ships features fast and leaves the codebase better than it found it.
Copy bootstrap/bootstrap-prompt.md and give it to your AI agent. The agent will:
- Analyze your project
- Adapt the templates to your stack
- Apply the framework
- Run the first quality check cycle
Grab what you need from templates/. Each file is fully annotated — every config decision is explained.
The paper explains the full framework with worked examples from real projects. See The paper.
The framework was developed and validated on open-garden-planner, an open-source CAD-precision garden planning desktop app (Python/PyQt6, GPLv3), and applied as a brownfield retrofit to a second, multi-stack project.
Real numbers from open-garden-planner:
- 247 commits over 99 days of AI-assisted development
- 79 releases shipped
- Full CI pipeline: lint + test + security on every push
The accompanying paper, Agentic Software Engineering: A Framework for Producing Maintainable Software with AI Agents, is published separately.
- Paper (PDF): to be added once finalized
- Preprint / DOI: to be added
- Sebastian Wienhold —
- Prof. Dr.-Ing. Mark Schutera — DHBW Ravensburg —
- OpenAI: Harness Engineering: Leveraging Codex in an Agent-First World (Ryan Lopopolo, 2026)
- Winters, T., Manshreck, T., & Wright, H.: Software Engineering at Google (O'Reilly, 2020)
- Birgitta Böckeler: Harness Engineering for Coding Agents (martinfowler.com, 2026)
MIT — use it, fork it, improve it, adapt it for your stack. See LICENSE.