Coherence

Most engineers use an AI coding agent like autocomplete and stop at convenience. Leverage comes from coherence: keeping intent, artifacts, and execution aligned so the AI works longer than you do.

Coherence means three alignments:

• Intent vs. artifacts — what you want and what is written down
• Artifacts vs. code — design and implementation
• Code vs. environment — what runs and where it runs

When a seam drifts, you land in a slow read–correct–retry loop.

Treat the AI agent like an employee, not a pair programmer. Employees need direction, expectations, memory, and feedback from the real world. I give those up front, then let it work asynchronously. A quick check: the agent should be working at least 2× longer than I’m chatting with it. If I keep intervening, I’m doing its job.

Fixing generated code is expensive; designing first is cheap. I’ll even ask one AI to draft the brief another AI will implement. That approach is cost control, not overengineering.

The biggest gains come from non-code artifacts that lock intent in place before code exists:

• Design docs and workflow notes
• AGENTS.md and other operating guides
• Failing tests that force alignment

The agent needs the same feedback loop I have. If I’m building a site, it should see the browser. If I’m building a CLI, it should run in a close match of the target environment. If I’m stuck manually validating everything, I’ve kept myself in the loop and throttled iteration.

Let the agent exercise the work and fill its own backlog: browse the site like a new reader, run the CLI in a staged environment, and capture TODOs for confusing flows or visual glitches without waiting on me to spot them.

Commits record outcomes, not dead ends. I start and end work by logging an issue in git-bug (or similar) so failed approaches and constraints live beside the code. Without that memory, the agent will “fix” problems by creating new ones forever.

The failure smell is simple: the agent keeps producing more work for me instead of less. When that happens, stop generating code, rebuild the artifacts, and restore alignment. Coherence keeps intent, artifacts, and execution in lockstep—humans set direction and taste; the agent supplies execution and endurance.