How many AI agents does a small business need? Count the roles, not the agents

Some vendors pitching AI to small businesses now quote prices per agent, as if agents were staff to hire. Anthropic, which sells the models that power those agents, takes a different line. Its own guide to building effective agents tells businesses that the right answer "might mean not building agentic systems at all."

The guide separates two things that sales pitches tend to blur. Workflows are systems where AI models and tools follow "predefined code paths". Agents are systems where the model "dynamically directs its own processes and tool usage". In plain terms: a workflow follows steps you set in advance, like a checklist; an agent decides its own next step. The guide's advice is to find "the simplest solution possible" and add complexity only when it's needed.

That changes the question small business owners ask. "How many AI agents do I need?" sounds like a headcount question, the way "I need a marketing team" does. But an agent isn't a person, and the work doesn't divide one-to-one. At The Modern Shrine, below, one agent running the whole social-media workflow hit a ceiling within about three weeks, while a single reviewer agent now checks the work of the others before it goes out. Counting agents the way you'd count staff tells you very little.

The questions that decide whether it works are different: what work needs doing, who owns each part, who checks it, and how it moves from one step to the next. Whether that ends up as one agent, one agent with five helpers, or a plain workflow with no agents at all is a design detail that comes last.

Start with roles, not a number

A role for an AI system needs the same things as a job description: the work it owns, the department it serves, the number it is meant to move, and who checks its output.

The Modern Shrine, a consumer AI product run as a one-person company, shows what that looks like. It runs on eight AI agents and no staff. That is one company's number, not a rule. Since 25 September 2026 each agent has had one role, written down like a job description:

• Chief of staff: hands out the work and decides what is finished.
• Growth pipeline: finds and qualifies leads.
• Research and knowledge: reads sources and keeps the notes current.
• Content and distribution: drafts posts and emails.
• Quality and safety reviewer: checks everything above before it goes out.
• Revenue operations: tracks deals and follow-ups.
• Product and automation: builds and fixes the tools the others use.
• Ops and insights: pulls the numbers into reports.

Two rules sit on top of the roles. The agent that writes something is never the one that reviews it, and nothing public goes out without the owner approving that exact batch. Keeping writer and reviewer separate is the main defence against invented details reaching the public. (Why one AI agent will confidently lie to you explains the reasoning.)

The team's social-reply workflow has published 15,621 replies, 98.7% of them live on the first pass, and the owner's routine daily review takes about 20 minutes. The method and the numbers are in the agent reply system case study.

The useful part of that list is not the eight. It is that every role has a written job and a named check.

The count hides the questions that matter

"How many agents?" crowds out the questions that decide quality:

• Which model does each job? At The Modern Shrine, as of August 2026, a cheaper, faster model handled the simpler work and a stronger one was kept for final checking.
• What rules surround it? Engineers call this the harness: what the AI may do on its own, what needs a person's approval, and when it must stop and ask.
• Which tools can it use? An agent with access to the CRM, the calendar and email can do very different work, and very different damage, from one that can only draft text.
• How does work pass between steps? Who hands what to whom, and in what order it gets checked.

The Modern Shrine's team got the shape wrong in both directions before it settled, and neither failure was about the count.

Too much in one place. In mid-2026, a single agent ran the whole social-media workflow. After about three weeks it hit a ceiling: too many invented details, and no way to track which step a task had reached. It was being asked to get two different things right at once, the technical details and the marketing.

No clear owner. After the team grew, a different failure showed up on 28 September 2026. The agents were busy commenting on, reviewing and editing each other's work, but very little was finishing. The fix was organisational, not technical: one domain, one owner, and one agent allowed to say "good enough, ship it".

In both cases, the change that helped was to roles and hand-offs: splitting the work in the first, and naming one owner in the second.

Sometimes the answer is no agents

Following Anthropic's distinction, many small business jobs are workflows, not agents. The steps are known in advance: read the enquiry, check the rules, rank the options, draft a reply, send it to a person to approve. Anthropic's guide describes workflows as offering "predictability and consistency for well-defined tasks": a chain of AI calls in a fixed order, with plain code for the hard rules.

In October 2026, a Flareworks AI audit of a Singapore plumbing, electrical and aircon firm with about ten technicians recommended exactly that shape for dispatch. Hard limits such as availability, skills and distance stay as plain rules in code, and the AI only ranks the two or three technicians who qualify. That is a proposal, not a result yet. The audit's estimate for about 100 dispatch conversations a month was US$3 to US$30 of AI usage, depending on the model.

One front door, not five tools

Splitting work across many agents has a cost that rarely shows up in a demo: every separate tool is another screen to learn, another login, another place to check.

In another October 2026 Flareworks audit, a 50-person home-repair group had already tried a team tool and dropped it, because technicians wouldn't log in. When a new dashboard was suggested, the owner's first question was "Who's gonna log in daily?" (The full audit is in Every AI proposal has a second spec.)

The same logic applies to agents. When several agents serve the same people, joining them into one workflow behind one front door, a chat channel or the tool the team already uses, avoids asking staff to learn and check several new tools. At The Modern Shrine, eight agents still come down to one daily review for the owner.

Anthropic's advice and these cases point the same way: start with the simplest system that does the job, define the roles first, and add an agent when a real job appears that nobody owns.

Originally posted on LinkedIn

View on LinkedIn ↗