I stopped asking my team to use AI. I asked them to manage it
One designer now does the work of three. Here's how a team stopped asking people to use AI and started asking them to manage it.
Why it matters
A practitioner case study on agent-team workflows: moving beyond individual AI productivity to multi-agent orchestration with human oversight, demonstrating measurable output gains (3x designer productivity, 1-day prototype cycles) and revealing the operational and cultural shifts required — role-based agents backed by subagents, human sign-off on all work, and the shift from task-doers to agent managers.
The key facts
11 to knowTeam structure: 7 people each work with a role-based primary agent plus specialized subagents; author uses agents for all functions
Productivity metric: each UX designer now produces what three used to; prototype cycle time reduced from weeks to ~1 day
Release velocity doubled; capability shipment measured weekly by function (requirements defined, interfaces designed)
Agent-to-agent collaboration: product management agents triage and research, design agents (Muffin, Ive, Sagmeister, Dieter) collaborate with frontend engineering agent (Travis); PacMan orchestrator routes work across functions
Cost: $2,000–$5,000 per person per month covering full role agent plus specialized subagents; cited as cheaper than single-developer AI coding benchmarks (Gartner: ~$20,000/month per developer)
Quality control: agents reach ~95% acceptance rate after feedback loops; code reviewed by separate agent in different context, then human sign-off required before PR; layered review (subagents check primary work, peer agents check engineering, final human pass)
Adoption timeline: took 'more time than expected' with 'a lot of micromanaging early on'; agents graduate to bigger work as they prove themselves; retrospectives after each project fold learnings into agent memory
Homegrown memory layer built early; later agents have native memory
Governance: Kanban board tracks agent progress; detailed control policies enforced by agent harness; all humans sign off before work moves forward
One-hour prototype delivery: FinOps dashboard went from whiteboard concept to live running prototype with real AI traffic in 1 hour with humans and agents working together
Lesson: early adopters moved fast; bottleneck shifted from execution to human review; now teaching agents to run 100-point review checklists automatically, leaving high-value assessment to humans
Go to the source
CIOcio.com
Publisher excerpt: My team was already using AI when I joined the company a year ago, and I quickly spotted a bottleneck. We’d finish large product requirements documents that then sat in inboxes for a day or two before someone read them and handed the work to an agent. To cut the cycle time, I had the recipient’s…