The first time you sit in a drum circle, you tense up.
No sheet music. No conductor. No agreed-on song. Just a pulse — and a circle of people you've never played with — listening, adjusting, finding the beat together.
By the tenth time, you've stopped looking for the chart. You're listening for the pocket.
That shift — from searching for structure to building it through attention — is the one I keep coming back to in how we design with AI right now. Improvisation isn't the absence of craft. It is the craft.
The industry talks about "orchestrating" multi-agent workflows like it's leading a jazz ensemble — everyone reading the same chart, waiting for the downbeat. That framing is comfortable. It's also wrong.
It's a drum circle.
Every agent you bring to the circle — Microsoft, Google, OpenAI, a specialized startup, your own internal models — shows up with its own rhythm. Its own strengths. Its own latency profile, failure modes, and tendency to drift if you stop listening. There is no canonical song. There is no shared chart. There is only the pulse you build between them.
The design problem isn't picking the right agents. It's defining the pocket they're all playing into.
Where I'm sitting with this most right now: designing a single dashboard that unifies the agentic tasks of both buyers and sellers on a marketplace.
Two sides. Two completely different sets of jobs to be done. And a circle of agents whose individual incentives don't naturally align:
→ A research agent surfacing seller leads, syncopating with an outreach agent drafting first-touch emails.
→ A negotiation agent on the buyer side, holding tempo with a fulfillment bot on the seller side.
→ A pricing agent and an inventory agent — two drums that have to lock in, or the entire rhythm collapses.
The work isn't the integration layer. The integration layer is the easy part — and treating it as the hard part is the most common mistake I see in agentic design today.
The actual work is three things, in this order:
→ Defining the pulse. What is the unifying tempo of this experience? What does one beat feel like to a user who doesn't know — and shouldn't have to know — that five agents are playing underneath?
→ Hearing the drift. Which agent is rushing? Which is dragging? Where does the rhythm break down at the seams between buyer-side and seller-side flows? Most multi-agent failures aren't capability failures — they're tempo failures.
→ Designing the listening. The interface is the conductor. Not the human. The dashboard has to surface the pulse, expose the drift, and give the user enough signal to adjust without turning every interaction into a debugging session.
When it works — when the agents lock in and the buyer experience and the seller experience feel like one song instead of two bands playing in adjacent rooms — the dashboard doesn't feel orchestrated. It feels alive.
That's the bar.
So I'll leave the question open: are we still looking for sheet music? Or are we getting better at designing the listening?
#MultiAgentSystems #AIOrchestration #ProductDesign #DesignLeadership #AgenticAI
