Make coding agents dependable in real software teams with the controls, checks, and operating practices needed for production
Key Features
- Turn CLAUDE.md, conventions, and definitions of done into an agent-ready repo contract
- Create guarded execution with scoped tools, sandboxes, hooks, and hardened devcontainers
- Build evals, telemetry, CI gates, and audit records for production use at team scale
Book Description
When teams move from one-off coding sessions to real software delivery, model quality is only part of the problem. The surrounding harness determines what the agent sees, what it can do, how its work is checked, and whether the result can be trusted. Harness Engineering with Claude Code begins with a clear model of what a coding-agent harness contains and how its parts work together. You then apply that model inside a real repository by shaping CLAUDE.md, conventions, Skills, tool scopes, permissions, sandboxes, and deterministic hooks. The next stage turns agent output into something a team can verify. You will combine software tests with adversarial checks and an internal regression set, instrument runs with OpenTelemetry, and track operational signals such as cost, latency, failures, and security findings. You will also connect agent-produced patches to GitHub Actions, SARIF gates, OIDC authentication, nightly regression policies, and reproducible audit bundles. The closing chapters focus on adoption beyond one developer. You will compare operating models for small teams, platform groups, enterprises, and regulated environments, then score your current setup with the book's maturity framework and use the result to plan what to improve next.
What you will learn
- Write a CLAUDE.md contract that changes agent behavior
- Set safe tool, permission, sandbox, and network boundaries
- Add hooks that block risky actions and enforce checks
- Build regression evals for your own repository
- Trace agent runs and monitor cost, latency, and failures
- Gate AI-generated patches through secure CI pipelines
- Package each run for reproducibility and audit
- Score harness maturity and plan the next stage of adoption
Who this book is for
This book is for senior software engineers, tech leads, staff engineers, engineering managers, and platform-team leads who already use Claude Code or another coding agent and want a more dependable way to run it on real codebases. It is especially useful for teams in regulated or compliance-heavy environments. Readers should be comfortable with Git, the terminal, automated testing, and CI/CD basics. It does not teach coding-agent fundamentals, and no AI or machine-learning research background is required.
