I stopped asking my team to use AI. I asked them to manage it

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 agent do the pre-read instead, sending questions back to the authoring agent as needed.

I realized we would see even more efficiencies if the agents interacted with each other the way human teams do. So, I stopped asking people to use AI to do their own jobs faster, and started asking them to hire and manage agents instead, like junior employees. They train them, set detailed expectations of outcomes, review plans, run periodic checks, make sure they collaborate with peer agents and own the quality of the output. A PwC survey of senior executives found the same thing: organizations adopting agents report gains, but the value concentrates where agents work across functions rather than in isolation.

Seven people on my team each work with a role-based primary agent, backed by subagents for specialized tasks, and I use agents for all of my functions. They’re full participants in the software development life cycle, not prototypes. Each person owns their agent budget and evaluates new tools for our stack. For example, the product management agents triage incoming customer requests, research and define requirements, and collaborate with peer agents.

Muffin, the product design agent, works in two modes: designing the frontend experience for a feature in the sprint, or exploring ideas not yet on the roadmap. In the background a separate agent responds to Slack alerts, triages issues and queues work for the project-manager and developer agents, pinging the person only when they are needed.

Muffin works with Travis, the frontend engineering agent, and delegates to her own subagents. A design system agent named Ive checks our components before anything new is built. Sagmeister reviews visual design and accessibility, and Dieter digs through old branches for context. A designer hands Muffin Figma mockups, and her team turns them into working code. The code goes to Travis, who runs it past his own subagents, so it meets our frontend engineering standards.

PacMan, an agent our research team built, sits on top as the orchestrator. It takes a ticket from Waffle, the product management agent, and routes the work: engineering to Travis, design to Muffin and questions about our proxy technology to Jarvis. A human signs off before anything moves forward. PacMan coordinates across product management, design, frontend and backend, and keeps the humans moving — maybe better than we managed on our own. “It’s unnerving that the agents are now driving the humans rather than the reverse,” says Jim Beno, our head of product design.

The moment I knew we had crossed a threshold came with a request for a new AI FinOps dashboard to show all the tokens consumed by each team, correlated to the work each team had done. Humans and agents worked in one room, and we went from a whiteboard concept to a functional prototype running live against real AI traffic in just one hour.

Agents earn trust the way junior hires do

Getting there took more time than I expected, and it required a lot of micromanaging early on. Agents graduate to bigger jobs as they prove themselves. Sometimes I handed one something too big, checked back 24 hours later, and the result was garbage. Over time they took on more complex work with less direction. After each project, I walked the agent through a retrospective, teaching it to notice what it had learned and fold those learnings back into its memory. Many agents now have more native memory, but ours was homegrown at the start.

When Muffin started delivering designs in code, they looked right on the surface but used the wrong components and missed our branding. Developers had to rework half the output, creating tension between teams. To resolve this, Alan Dong, the engineer who built Travis, loaded frontend best practices into its knowledge base. Separately, we changed how we prompt: treating the agents like junior hires, giving them more context than feels necessary, and debriefing on lessons after each session. We built a workflow where Muffin creates the designs in code and Travis reviews them. That collaboration got us to roughly 95% acceptance from the frontend team.

Another time, a wording mix-up nearly triggered a premature build. Katie Chen, senior product designer, selected “finish the load animations” from a list of options, just to point Muffin at the task. Muffin took it as a green light and started building. Katie stopped her mid-stream. Now Muffin asks clarifying questions first, and nothing gets built without an explicit go-ahead.

Early agent adopters moved so fast that the rest of the organization couldn’t keep pace. One person on the team built a full-featured Model Context Protocol (MCP) catalog that sat in engineering for months before anyone built on it. We’ve since automated the path from specification to ship, which only moved the constraint. The human reviewers are now the bottleneck.

So, we’re teaching agents to out-review humans. Different people ask different review questions, so we aggregate them into a checklist the agents run every time. Then we ask Claude for other industry review best practices and teach those too. Eventually, an agent reviews things 100 ways automatically, leaving less that requires a human eye. A well-trained agent carrying the load, with humans doing the high-value assessment, beats what a human could do alone. No one has time for those 100 checks plus their own advanced 10.

A human still signs off on everything

For oversight, we use a Kanban board the team has run for years to track progress. Treat agents like team members and the rest follows: prioritize their work, scope it clearly, check progress, review output and iterate. Code from one agent gets reviewed by another agent running in a separate context window on a different underlying model. A pull request triggers another round, and the agents can spend an hour iterating until it’s clean. No agent opens a pull request without a human signing off first. The review is layered: subagents check Muffin’s work, Travis’s team checks engineering and the pull request gets a final pass. A detailed set of control policies, enforced by the agent harness, holds the security line — all written down and built into the agents’ instructions, so they know when and how to ask for human review.

One designer now does the work of three

Going from idea to prototype used to take a few weeks. Now, it takes about one day. We built new in-product dashboards measuring the value of AI work in 25 hours, back before our review processes were in place. Release velocity has doubled. We measure output weekly, with a metric that fits each function. For product management, it’s requirements defined; for UX, it’s interfaces designed. At the team level, we track the total capabilities we ship per period across every function. By that measure, each UX designer now produces what three used to. We also respond faster to feature requests, information requests and asks from sales enablement and marketing. 

This is also cheaper than what other companies are reporting. Most token-cost benchmarks measure a single developer writing code. My team spends $2,000 to $5,000 per person per month, covering a full role agent, plus specialized subagents for product, design and marketing — not just engineering. Measured against what that replaces, and against per-developer AI coding costs that Gartner says are on track to rival monthly salaries — with reports of single developers consuming $20,000 in a month — it’s reasonable. We’re deliberate about cost, routing simple prompts to cheaper models, and running small models on laptops where tokens cost us nothing. Together, these gains have let us defer hiring new staff.

For leaders who want to try this, here are three pieces of advice:

1. Study what peers are doing in any organization. There are AI black belts publishing best practices. Learn from them, then lead your team.
2. Push your teams to collaborate early. It keeps work aligned with production standards, so projects don’t have to be rebuilt, and produces higher quality faster.
3. Don’t start from scratch. The space moves fast enough that plenty of tools already exist to build on. You’ll still need to shape them to how your team works, but you don’t have to be an expert to start.

The bigger change has been cultural. People stay experts in their own lanes, but that expertise now guides and refines work others used to own outright. As Katie puts it: “The net effect is we all end up acting more like product owners, not just holders of one job title.”

In the end, everyone becomes a manager of agents — in any role.