“11 AI agents” is an empty metric

Agent count is not an architecture explanation

Agent count can look impressive on a slide. Simon Willison's quoted line from Boris Mann cuts through that: saying "we have 11 AI agents" is about as informative as saying "we have 11 spreadsheets" or "we have 11 browser tabs". The number does not describe the work, the quality or the responsibility model.

The word agent covers many different things

Willison shared a short observation about the language around agentic systems. Mann's point is not that more agents are always bad. The critique is aimed at a metric that replaces an explanation of architecture with a simple count.

That matters because the word agent is being used for many different things: a standalone tool, a planner, a model wrapper, an automation step or a team of specialized processes. Without context, the number does not say what the system actually does.

The agent count hides the questions that matter

Agent marketing often sells complexity as proof of maturity. But a buyer or user does not primarily need to know how many internal parts a system has. They need to know what task it completes, where it stops, what it is allowed to do on its own and when it must ask a human.

More useful measures are outcome, responsibility boundaries, handoff quality, runtime visibility, failure handling, permissions and human review. Those are the signals that separate a working workflow from a diagram with impressive labels.

Instead of asking "how many agents do you have", ask who decides, who checks, who is accountable for the output and what happens when something fails. A useful agentic system has clear roles and clear limits.

Good questions include: how is context handed off, how are steps logged, how are conflicting instructions handled, which actions require approval and how easy is it to inspect or roll back the result. Without those answers, agent count is just decoration.

Questions that a headcount will never answer

Multiple agents can make sense when they separate planning, execution, review and testing. They can also add latency, cost and new failure points. The count alone does not distinguish careful orchestration from chaos.

That is why the short quote works as a useful test for any agent pitch: can the vendor explain the outcome and the control mechanisms without hiding behind the number of agents?

For new agent products, observability, permissions, audit trails and the quality of human review will matter more than the headline count. That is where we will see whether agents are helping the work or just adding another abstraction layer.

If a presentation starts with the number of agents, that is not automatically a problem. The problem starts when the presentation also ends there.

Lilith's verdict

A useful antidote to agent marketing that treats count as sophistication. Eleven agents can be a deliberate system, or eleven places where context can get lost. Without clear boundaries, auditability, permissions and human review, it is closer to an inventory than an architecture.