# The Essence of AI Coding Is Not the Model but the Harness: Matt Pocock’s Agentic Engineering
When people discuss AI coding, they usually name models first: Claude, Codex, Gemini CLI. Models matter, but Matt Pocock points elsewhere. The real difference comes from the harness.
The harness is the working environment around the model: prompts, skills, codebase structure, tests, documentation, sandboxes, GitHub Actions, and review flow. It is like judging not only the engine of a car but also the chassis, pit crew, and track operations.
This matters because model performance is hard for us to control, while the harness is something we can design.
## AI has eaten tactical programming
Pocock borrows John Ousterhout’s distinction between tactical and strategic programming. Tactical work includes writing code, fixing bugs, making commits, and matching syntax. Strategic work is deciding what structure will be maintainable, how to divide work, and where the codebase should go.
AI has already absorbed much tactical programming: small features, tests, refactoring drafts, and documentation updates. As AI handles more tactics, human value moves toward strategy: clear goals, narrow scope, completion criteria, and tests.
## The work environment matters more than the latest model
The strongest line is that everyone obsesses over models, but we should care more about the harness. For example, if you want to reduce token cost, one answer is shorter prompts. Matt’s answer is a codebase that is easy to change. Clear structure, tests, and current documentation let AI work with less context. A tangled codebase makes even expensive models wander.
## Skills should be managed as procedures, not piled on
A skill is a reusable bundle of instructions for repeated thinking or work. It can make an AI act as a learning coach, critique a design aggressively, or review PRs in a specific way.
But Matt does not say to add as many skills as possible. He suggests deleting skills, plugins, MCP servers, Claude.md, and agents.md, then observing the model in a blank state. Add back only what is truly needed. Too many instructions can pollute the context window.
## AFK agents are closer to queues than infinite loops
“Agentic loop” sounds attractive: the agent thinks, acts, observes, and acts again. In practice, it can blur scope, raise cost, and remove review points.
Matt proposes “queue rather than loop.” Put work into a queue, like GitHub issues or Jira tickets. The agent takes one task, investigates, changes, tests, makes a PR, and a human reviews it. Good candidates include failing-test investigation, README updates, refactoring proposals, PR review drafts, security checklists, and old-issue reproduction.
## AX: Agent Experience now needs design
Developer Experience made environments easy for humans to install, run, test, and deploy. Agent Experience is the degree to which an AI agent can work in the codebase.
- Predictable folder structure
- Clear test commands
- Automated type checks and linting
- Current README and development docs
- Clear module boundaries
- Safe validation of small changes
- Enough information to run in a sandbox
Good AX overlaps strongly with good DX. What humans can infer informally, agents often miss; therefore documentation, tests, commands, and boundaries matter more.
## Problems found by AI should become system improvements
If a model finds a security bug, do not stop at “this model is good.” Ask why the bug remained, why tests missed it, whether similar bugs exist, and how future checks can be automated.
AI output should be a signal to improve the harness: add tests, revise review criteria, create a security-check skill, or add CI checks. That turns AI coding from a one-off productivity tool into an organizational learning system.
## Product and business judgment still belongs to humans
The video also touches on SaaS and AI startups. Matt’s answer is simple: talk to customers, find real problems, prototype, and validate. AI accelerates implementation, but it does not decide what to build, why to build it, or what to remove.
## Seven things Korean developers and teams can do now
- Read your README as if an agent were entering the repo for the first time.
- Turn repeated requests into skills or templates, but only when they are truly repeated.
- Split issues into AI-sized tasks.
- Include test commands and done criteria in instructions.
- Start AFK work inside sandboxes and limited permissions.
- Review AI PRs for failure patterns, not only code.
- Follow model news, but check structure, tests, documentation, and review flow more often.
## Related reading
- The agentic AI era
- Training AI agents with Hermes Agent
- Claude Skills for small business
- Frontend basics for vibe-coding beginners
- Headroom and AI token diets
## FAQ
### What is a harness in AI coding?
The full environment in which the model works: prompts, skills, code structure, tests, docs, sandbox, CI, and review flow.
### Why is codebase structure more important than the latest model?
A clear structure and tests let AI make safe changes with less context. A messy codebase makes any model struggle.
### How is a queue different from an agentic loop?
A queue lets humans define tasks and review results one by one. A loop can become open-ended and harder to control.
### What is AX?
Agent Experience: how easy it is for AI agents to work in a repository. It overlaps with DX but raises the standard for clarity.
### What should be prepared first before assigning coding to AI?
Scope, completion criteria, tests, sandboxing, permission limits, and review flow.
## References
- Tech Bridge video on Matt Pocock’s agentic engineering workflow
- Matt Pocock skills repository
- Matt Pocock on X/Twitter
AI coding’s next step may not be turning on more tools. It may be pausing to inspect the environment where AI works: missing docs, fragile tests, and tasks that can be queued. That small cleanup can matter more than subscribing to one more model.