From an AI-Using Team to an AI-Managing Team
1. Manage agents like new hires
Assign work, explain expected outcomes, give feedback, and expand their scope.
- Start with small tasks at first and check progress frequently.
- After a project ends, reflect on what was learned and apply it to the agent’s memory and work instructions.
- People set direction based on their expertise and are responsible for the quality of the output.
2. Collaborate through role division and cross-review
Specialized agents hand work off to one another and review each other’s results.
- Waffle defines the requirements, and PacMan assigns work to the appropriate agent.
- Muffin handles product design, and Travis handles frontend implementation and review.
- Sub-agents check the design system, accessibility, and engineering standards.
- Automate repetitive review items, but keep human approval for moving to the next stage and creating PRs.
3. Measure results by both workload and cost
This team measured effectiveness based on role-specific outputs and the features actually launched.
- The time from idea to prototype was reduced from several weeks to about one day.
- The UX designer’s output was reported to be equivalent to three people, and launch speed became twice as fast.
- The cost of role-specific and specialized agents was $2,000 to $5,000 per person per month, and productivity gains made it possible to postpone additional hiring.
