You Are Part of the Harness

AI agents amplify your habits. All of them.

I spent six posts explaining how to constrain AI agents. The biggest constraint was on me. I skipped planning. I believed in DRY but never enforced it with tooling. I knew principles mattered but never wrote them down. These were bad habits I could tolerate when I was the only one writing code. With 100+ agents, those habits scaled too.

The harness does not just constrain agents. It forces you to curb the bad habits you have been getting away with for years.

From your first coding class, the instructor tells you to plan before you write code. Pseudocode first. Think through the logic. Then implement. It is so easy to ignore. I ignored it for years and got away with it.

Agents made that impossible. My orchestrator used to carry the full plan in its context window while also executing. By the third delegation, it was already forgetting the first. Planning and execution were tangled together, and the system was falling apart.

So I separated them. Planning agents decompose the work upfront. Every task gets filed as a git-synced issue before any code gets written. Then execution agents pick up those issues independently. A planning version of the system and a coding version of the system. Exactly what my instructor wanted from me twenty years ago.

The hard part was not building that separation. It was admitting that I had been skipping the planning step my entire career and getting away with it because I could hold it all in my head. The agents could not.

The harness starts with you. It depends on you articulating your engineering principles. Being clear on your architecture, on what you value, so that your agents value it too. If you have never written down why you structure code the way you do, your agents will never know. They will fill in the gaps with whatever gets the tests to pass.

The difference between vibe coding and agentic engineering was never the tools. It was always whether you were willing to do that work.