The Question Nobody Is Asking at the Right Time
Most non-tech operators are being sold AI agents before they've asked a foundational question: should this be one persistent agent working continuously toward a goal, or a swarm of specialized agents coordinating in parallel?
Ethan Mollick published a sharp frame for exactly this choice this week, timing it around OpenAI's DevDay release of 'Dots' — always-on GPT-6 agents with their own cloud compute running 24/7. The distinction he draws isn't academic. It determines how you structure oversight, who's accountable when something goes wrong, and what your implementation actually costs in time and coordination overhead.
What the Distinction Actually Means for Operators
A single persistent agent (the 'dot') is easier to supervise. You can trace its decisions, audit its outputs, and assign clear ownership. The tradeoff: it's sequential. It works through tasks one at a time. For most operators just beginning to deploy agents in workflows — sales follow-up, contract review, customer triage — this is the right starting point. Lower risk, faster trust-building, cleaner accountability.
A swarm unlocks parallelism. Multiple agents divide labor, hand off tasks, and converge on an output. That's genuinely powerful for complex research, multi-step pipeline orchestration, or anything that benefits from concurrent processing. The cost: coordination risk multiplies. Agents can contradict each other, loop, or produce outputs no single human reviewed in sequence. Swarm architecture requires governance infrastructure most non-tech orgs don't have yet.
With sub-2% of people currently using agents at any meaningful depth (per Allie K. Miller), most operators are choosing architecture before they've built the operational muscle to manage either model well.
The Actionable Takeaway
Before your next vendor pitch — and DevDay guarantees there will be one — run every proposed agent deployment through two questions:
- Who reviews this agent's output before it affects a customer or a system of record? If you can't answer that clearly, a swarm will make it worse, not better.
- Does parallelism actually help here, or does it just sound faster? Most GTM and ops workflows are sequential by nature. A well-configured single agent with clear handoff points beats a swarm you can't audit.
Start with the dot. Earn the right to the swarm by proving your oversight process works at the simpler level first. Vendors will pitch swarms because they're impressive. Your job is to ask whether your organization is ready to govern one.