A working draft from my notebook. The copy is provisional and will be reviewed before public launch.

Start with the work, not the agents

An AI development team sounds simple when described as three roles: someone plans, someone designs, and someone builds. But those roles only become useful when they share a clear understanding of the result.

The question I’m exploring is how to make that coordination practical. I want a workflow where each specialist can move the work forward without losing the intent of the original brief.

Give every role a boundary

A planning agent needs to understand scope and dependencies. A design agent needs context about the person using the product. A developer agent needs requirements that are concrete enough to implement.

Giving every agent the same broad instruction is not the same as giving the team a shared goal. Responsibilities need to be explicit, including the point where a role hands the work to someone else.

The handoff is part of the product.

Make human review a real step

For my initial setup, approval is a deliberate part of the workflow. A person should be able to inspect the proposed direction before an important action becomes difficult to reverse.

The review needs evidence: a visible design, a working implementation, or a specific explanation of what changed. “Done” on its own does not tell you enough.

WORKFLOW NOTE
A shared brief
      ↓
Plan → Design → Build
      ↓
Human review + verification
      ↓
The next iteration

The questions I’m still testing

How much context should be shared? When should an agent stop and ask? What does a useful verification step look like?

I’m treating these as engineering and product questions rather than assuming the architecture has already solved them. The aim is a workflow that helps someone make something real.

✳

Thanks for thinking along.

Have a different perspective? Let’s talk