Aug 08, 20261 min read

Building Better AI Workflows

A practical look at structuring research, implementation, testing, and verification when working with coding agents.

AISoftware Engineering

Good work with coding agents is less about finding a single perfect prompt and more about creating a reliable loop around the model. The useful unit is a small, observable workflow: understand the context, make a focused change, and verify what happened.

Make the work legible

Before implementation, write down the shape of the problem. A short plan gives the agent boundaries, exposes missing decisions, and makes it easier to tell whether a result is complete. The plan should be specific enough to guide the work without turning into a second implementation.

typescript

type WorkflowStep = {
  name: string
  run: () => Promise<void>
}

export async function execute(steps: WorkflowStep[]) {
  for (const step of steps) {
    await step.run()
  }
}

Diagram of the understand, implement, verify loop feeding back into itself — The loop repeats until the result is verified, not just produced.

Close the loop with verification

Verification is not optional — it's the step that turns a guess into a fact. At minimum, that means running type checks and the relevant test suite, the same habits described in The Shape of a Useful Code Review.

Type checks pass without new suppressions
Linting is clean, not silenced
The relevant test suite runs and passes
A manual pass through the changed surface, not just the diff

A minimal loop

Read the diff before running anything
Run the checks that actually cover the change
Confirm the result matches the original intent, not just that it compiles

A workflow you can't observe is a workflow you can't trust.

Checklist showing understand, implement, and verify as sequential steps — A minimal checklist is often enough to keep the loop honest.

None of this is exotic. It's the same discipline that makes any engineering work reviewable — applied consistently enough that it holds up when the agent, not just a human, is the one making the change.