AI & automationSmall-business IT

One AI Agent or Many? Why Anthropic Says Simpler Business Bots Often Work Better

Anthropic says one AI agent with skills often beats a team of subagents on quality, cost and speed. What that means for a small business building an AI helper.

By Cybercave Technologies, LLC · Oct 27, 2026

A tidy white desk with a laptop, an open notebook with a handwritten checklist and a pink sticky note reading DONE, colorful sticky notes, a phone and two small potted plants — One clear checklist often beats a pile of handoffs, for people and for AI agents. Credit: Photo by Jakub Zerdzicki via Pexels (Pexels License).

For most business uses, one AI agent that loads the right instructions and tools when it needs them works better than a "team" of specialist AI agents passing work to each other. That's the advice in a new guide from Anthropic, the company behind Claude, which says the single-agent design often wins on quality and frequently costs less and runs faster [1][2].

What to do if you're building or buying an AI helper:

Start with one agent, one clear set of instructions and a handful of simple tools.
Write 10 real customer questions with the right answers, and test the agent against them before and after every change [1].
Split work into extra agents only when you can measure a problem that one agent can't handle [1][4].
Ask any vendor how many separate agents sit behind their product and how they test the handoffs between them.

Not sponsored.

What did Anthropic actually recommend?

Anthropic says many companies would get better results by folding their specialist "subagents" back into one main agent. A subagent is a separate AI worker with its own instructions, tools and memory that a coordinating agent hands tasks to [1].

According to the guide, Anthropic's Applied AI team compared designs across several large business deployments. A single agent that pulls in "skills" (short instruction packs it loads only when needed) often beat both one giant all-purpose prompt and a multi-agent setup on quality, and often cost less per task and answered faster [1]. In one comparison a customer ran, the single-agent version finished in roughly half the time [1][2].

Anthropic also released a free starter kit on GitHub to help teams test the switch. It works in three steps: measure the current system, merge agents that always run together, then fold the rest in one at a time and keep any agent whose replacement scores worse [3].

Why would one agent beat a team of agents?

Mostly because every handoff loses information. When one agent passes a summary to another, details drop out, a problem Anthropic compares to a game of telephone [1].

The guide lists other problems that pile up when each department builds its own agent: overlapping agents with no clear owner, a shared prompt full of rules that contradict each other, and extra wait time for customers [1]. The Decoder, which reported on the guide Oct. 25, noted Anthropic's example of one company whose dozen specialist subagents made it hard to keep its customer-facing assistant consistent and on-brand [1][2].

Anthropic says newer AI models can now plan and check their own work inside one loop, so much of the routing structure teams built around older models is no longer needed. It says it cut the system prompt for its own Claude Code tool by more than 80% on newer models, partly by loading skills and tools only when they're needed [1].

When do multiple agents still make sense?

Anthropic isn't saying subagents are dead. Its guide names three cases where a separate agent still earns its place [1][2]:

One agent's performance measurably drops because it's carrying too much information at once.
Several tasks need to run at the same time, each with a fresh, throwaway workspace.
One agent keeps picking the wrong tool, or its instructions conflict.

OpenAI's own guide for businesses lands in a similar place. It recommends getting the most out of a single agent first, and splitting into more agents only when instructions get too tangled or the agent keeps choosing the wrong tool among similar ones [4].

How many tools can one agent handle?

More than you might think, as long as the tools are clearly different. Anthropic says an agent's tool choices tend to get worse once it's given more than about 30 to 50 tool definitions at once, which is why it suggests grouping tools by what the customer is trying to do [1]. OpenAI says some systems handle more than 15 well-described, distinct tools without trouble, while others struggle with fewer than 10 that overlap [4].

For a small business, that's usually plenty. Looking at many overbuilt agent designs it has reviewed, Anthropic writes that one clear set of instructions, four or five simple data tools and firm rules on what the agent may say would have covered most of the job [1].

What's the catch?

Merging agents can quietly break things, so test before you trust it. In one Anthropic test with a sample support agent, folding one agent into another made it faster but broke a step that had been working: the main agent started answering a delivery-date question itself instead of checking with the shipping specialist [1][2].

That failure showed up in only one of two test runs, which is why Anthropic advises running each test more than once and making sure your tests cover every rule the agent must never break [1].

What does this mean for a small business?

Simpler is usually cheaper and easier to keep right. If you're adding an AI helper for things like answering common questions or booking appointments, one well-tested agent with clear limits is a sensible starting point, and the extra layers can wait until a real, measured need shows up [1][4].

It also gives you a good question for vendors: if a product advertises a "team" of AI agents, ask what each one does, how answers are checked when one hands off to another, and how they test for mistakes. For more on keeping AI costs and risks in check, see our pieces on the hidden second bill for AI seats, setting a spend cap on AI keys and trying an AI assistant carefully.

FAQ

What is an AI "subagent"?

It's a separate AI worker with its own instructions, tools and memory that another agent hands tasks to, then gets a result back from [1].

Is Anthropic's migration kit free?

The kit is posted publicly on GitHub under an open-source license, but Anthropic labels it a reference implementation that isn't maintained and doesn't take contributions. Running the tests it sets up still uses paid AI model time, and the kit stops to show a token estimate before any paid run [3].

Do I need a test set before changing my AI setup?

Anthropic recommends it, and says even 10 "golden" example questions with known right answers are enough to start [1].

Does this apply to tools other than Claude?

The guide is written for Anthropic's platform, but OpenAI gives similar advice for its own agent tools: get the most out of one agent before adding more [1][4].

This article is general information, not technical, legal or investment advice. Mentions of companies are not endorsements.