Agent Harness Engineering

Description

As AI agents become increasingly autonomous, better prompts alone are no longer enough. Prompts are suggestions agents can ignore or forget between sessions. Reliable AI systems are built through strong harnesses—the infrastructure layer that wraps an AI agent, enforces behavior, and validates results.

This training solves three real problems: agents claiming victory too early even when tests fail, losing context and repeating work between sessions, and building more than was asked because scope is undefined.

A harness has five subsystems working together:

• Instructions — What the agent should do, and in what order (CLAUDE.md, ADRs, domain language, progressive disclosure)
• State — What's been done, what's in progress, what's next (progress logs, feature lists, git history, session continuity)
• Verification — Only passing tests count as evidence (linting, mutation testing, E2E runs, no early victories)
• Scope — One feature at a time, no overreach (feature lists, explicit done definitions, constraint enforcement)
• Lifecycle — Initialize in an isolated sandbox before work, clean up after (health checks, file boundaries, permission controls, progress handoff, recovery paths for next session)

In this hands-on course, you learn how to design, build, and operate these five subsystems using a two-directional control model:

• Guides — Feedforward controls that anticipate unwanted behavior before the agent acts. Example: an ADR the agent reads first; a skill list that removes unreliable tools.
• Sensors — Feedback mechanisms that detect and correct issues after execution. Example: a hook that runs tests on every commit; Semgrep rules that catch forbidden patterns.

You'll explore how these work together to steer and verify agent behavior across multiple sessions, preventing common failures: agents claiming victory too early, losing context between sessions, building more than was asked.

You already use skills, subagents, and hooks in your agent workflows. This training turns these separate pieces into one integrated system—a harness that governs every phase of agent execution.

Using practical examples, you'll build a complete harness around an existing codebase. You'll work with concepts such as CLAUDE.md hierarchies, Architecture Decision Records (ADRs), domain language, linting, Semgrep, mutation testing, context partitioning, observability, and session handoff patterns.

By the end of the course, you'll be able to design a complete agent harness that makes agents work reliably within your team, with errors caught before they reach production.

Prior Knowledge

You should have:

• Hands-on experience using an agentic coding tool (Claude Code, Copilot agent mode, Cursor) on a real codebase.
• Familiarity with instruction files (CLAUDE.md or AGENTS.md) and what they contain.
• Basic understanding of skills, subagents, and hooks—how they steer and constrain agent behavior.
• Knowledge of sessions and context windows (what's lost between runs, what persists).
• Comfort with git (branches, commits, history) and automated testing/linting.

This is not a first-time AI-assistant course. You need to have used agents before to get value from this material.