Agent = model + harness — get AI coding agents to production standard
An advanced course on harness engineering: what it takes to put AI coding agents like Claude Code and Codex into real work reliably. You design the harness — context, tools, verification, automation loops — rather than the model, and turn the agent into a development partner you can trust.
What is harness engineering?
Harness engineering is the skill of designing the infrastructure and control patterns that make AI coding agents reliable enough for production. The industry frames it as agent = model + harness: what determines an agent's reliability is not the model but the harness around it — context, tools, permissions, and verification loops. Through 2026, Claude Code and Codex CLI have moved past early adoption, multi-agent workflows have moved from demo to production, and MCP has settled in as the connection standard. This course is not about trying the agents out; it is about taming them with autonomous cycles and verification gates until they work as development partners.
Harness first
Reliability comes from designing context, tools, and verification loops, not from swapping models.
Autonomous cycles
The Plan-Work-Review loop lets the agent plan, execute, and review its own work.
Production oriented
Observability, auditing, security, and governance are covered, so the result can actually be deployed.
Evidence and cases
Why this training matters now
Stripe: deployed to 1,370 engineers company-wide
Stripe rolled out Claude Code to 1,370 engineers at every level using a zero-config enterprise binary. One team finished a 10,000-line Scala-to-Java migration in four days, work that had been estimated at 10 engineer-weeks, or about two and a half months. It is the clearest example of a harnessed agent accelerating a large migration.
Source: MindStudio, 'Codex vs Claude Code 2026'
Anthropic: 67% more pull requests merged
After adopting Claude Code across its engineering organization, Anthropic reported roughly 67% more pull requests merged per engineer per day. Staff self-reported an average productivity gain of 50%, and 14% of respondents were power users who reported gains above 100%. This is evidence that agentic coding delivers measured productivity, not just demos.
Source: Anthropic, 'How AI Is Transforming Work at Anthropic'
Curriculum
16 hours (2 days) — what's covered
What harness engineering is
4HAgent = model + harness, the harness as the thing that determines agent reliability, and Claude Code compared with Codex
Designing context and tools
4HContext design in CLAUDE.md and AGENTS.md, connecting MCP tools, and setting permission and file boundaries
Autonomous cycles and verification
4HThe Plan-Work-Review loop, test and verification gates, and patterns for running multiple agents in parallel
Governance and operations
4HTeam rollout strategy, observability and audit logs, security and data protection, and a production readiness checklist
Key concepts in the curriculum
A quick look at what this course covers.
The harness that tames the model
Binding a powerful model into a controllable system
Autonomous cycle
An agent loop that runs plan, work, review on its own
Agents in parallel
Several agents building together at once
Learning objectives
- Understand how AI coding agents work and what a harness is
- Design context, tools, and permissions for Claude Code and Codex
- Build the Plan-Work-Review autonomous cycle with verification gates
- Establish team-level agent governance, observability, and security
What you take away
- A CLAUDE.md and AGENTS.md context design for your own codebase
- One MCP tool integration with permission boundaries configured
- A Plan-Work-Review workflow with verification gates
- A governance and security checklist for team agent adoption
Expected outcomes
- AI coding agents you can rely on in production
- A structural lift in pull request throughput and delivery speed
- Large migrations and repetitive work automated
- Safe, team-level governance around agent use
How teams put this to use
Large-scale code migration
Accelerate legacy conversion and language migration with agent loops
Autonomous pull requests
Build a workflow that plans, implements, tests, and reviews from an issue on its own
Parallel multi-agent development
Split independent work across several agents and develop features in parallel
Shared team context
Bundle team conventions and tools into context so agent output stays consistent
Harness engineering for AI coding agents — FAQ
Tool training teaches you which commands to type. Harness engineering teaches you how to make an agent trustworthy — context design, permission boundaries, verification gates, autonomous cycles. That production reliability is the decisive difference.
The course targets developers who read and write production code. Comfort with Git and the terminal plus experience working in a team codebase is enough. It suits tech leads and platform engineers particularly well.
The course centers on harness concepts that are not tied to one tool, and covers Claude Code and Codex side by side. You learn the strengths and differences of both, and design principles that apply whichever you use.
Yes. Bring your codebase, or a representative sample, and you will build the matching CLAUDE.md and AGENTS.md context, tool integrations, and verification gates — ready to use right after the course.
The final session gives real weight to permission boundaries, observability and audit logs, sensitive data handling, and governance. You design the production security patterns that control what an agent can reach and what it can change.
Harness engineering for AI coding agents — get started
Tell us your goals and headcount, and we'll design the program around them.