How many AI agents does a business need?

The short answer

Most businesses need between one and six. One to start, and a lead plus three to five specialists once it matures. Organisations running dozens usually have a handful doing real work and a long tail that was created for a reason nobody remembers.

That is not a limitation of the technology. It is a property of how much distinct work a business actually has.

Why count is a vanity metric

Three reasons it misleads:

• Agents are cheap to create and expensive to maintain. Making one takes a minute. Keeping its instructions current, its connections alive and its output reviewed is ongoing, and it is the part nobody counts.
• More agents means more coordination, not more capacity. Every additional agent adds handoffs that can fail, permissions to manage and a memory that knows only part of the picture.
• An unused agent is worse than no agent. It has stale instructions and live credentials. When somebody eventually routes work to it, it does something out of date with real access.

The rule

A distinct job justifies a distinct agent. A large job does not.

Distinct means it has its own success criteria, its own tools, and would sensibly have its own memory. Large just means there is a lot of it — and a lot of one job is still one job, handled by one agent with a subagent or two for the steps that crowd its context.

The business-card test: write the agent's purpose in one line without using "and". If you cannot, you are describing two agents, or a subagent of one that already exists.

What a realistic team looks like

For a small company running this properly, roughly:

• A lead you talk to, which delegates. Worth having as soon as there are more than two specialists, because otherwise you are the router.
• A front desk on the channels customers use.
• An operations agent for inbox, coordination and the movement of data between systems.
• A research agent for monitoring, competitors, prospects, whatever recurring reading matters.
• One or two narrow specialists tied to something specific — a system only they connect to, a domain with its own rules.

Plus subagents inside those, which do not count against the number in any meaningful way because they add no supervision surface.

Signs you have too many

• You cannot name what each one does without looking.
• Two of them would answer the same question differently.
• Some have not run in a month.
• You find yourself deciding which agent to ask. That means the lead is missing, or the boundaries are unclear.
• The org chart has nodes that never light up.

Signs you have too few

• One agent's instructions contradict themselves.
• Its context is full of unrelated work.
• You cannot tell which part of a long prompt caused a bad answer.

Those are the three signals from multi-agent vs single agent, and they are the only reliable reasons to add one.

What about the catalogues of hundreds?

Platforms — ours included — ship large libraries of ready-made specialists, and it is worth being clear about what they are for. A catalogue is a starting point, not a shopping list: you install the two or three that match jobs you actually have. The value is not having 180 agents available, it is not having to write the one you need from scratch.

Same for starter teams. Installing a team of five and cutting it to three is faster than designing three from nothing, and cutting is the step people skip.

The number that matters instead

Not how many agents you have — how many distinct jobs are being done without you. A business with two agents doing two jobs reliably is further ahead than one with twenty agents and a person still checking everything.