One AI agent with skills often beats a team of bots. Start there.

Many owners picture their first AI rollout as a small digital staff: one agent for sales, one for support, one for invoices, and a manager agent to keep them in line. It looks great on a whiteboard. Anthropic's Applied AI team now says that, in practice, it often works worse.

What Anthropic found

In new guidance on building agents, Anthropic's Applied AI team compared designs across several enterprise deployments. Their finding: "a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task."

A skill here is a packaged set of instructions and resources the agent loads only when a task calls for it, instead of carrying every rule in one giant prompt. One customer that compared its multi-agent setup with a single agent loop saw "a roughly 2x faster runtime" with the single agent.

Anthropic also names a familiar trap: the agent design "becoming a copy of their org chart." Each handoff between agents is a place where context gets lost, rules get applied twice, or nobody owns the final answer.

Why this matters for a small business

Most small businesses are not building agent platforms. They are wiring an assistant into email, a CRM, a help desk or the books. The lesson still applies, maybe more so:

1. 01

   Fewer moving parts.

   One agent is easier to test, explain to staff and switch off if something goes wrong.

2. 02

   Lower spend.

   Every extra agent is extra model calls. Anthropic reports the single-agent design was often cheaper per task.

3. 03

   Clearer ownership.

   When one agent did the work, you know where to look when an answer is wrong.

When a second agent earns its place

Anthropic does not say subagents are always wrong. The guide keeps them for three situations:

1. 01

   1.

   The agent's context is getting so long that quality drops.

2. 02

   2.

   Pieces of the work can truly run in parallel.

3. 03

   3.

   The agent struggles to choose between tools or follows conflicting instructions.

On tools, the guide notes that once an agent has more than about 30 to 50 tool definitions, its choice of tool tends to get worse. If you find yourself there, split the work. Until then, one agent is usually the better bet.

Five steps for your first agent

1. 01

   1. Write 10 test tasks.

   Pick real requests from the last month and write down the right answer for each. Anthropic says even 10 golden examples is enough to start an evaluation.

2. 02

   2. Give the agent a few tools, grouped by job.

   In several reviews, Anthropic's team found that one clear system prompt, 4 to 5 thin data tools and firm output guardrails would cover most of the use case.

3. 03

   3. Put the hard limits inside the tools.

   Approvals for anything that sends, pays or deletes belong in the tool itself, not in a polite line in the prompt.

4. 04

   4. Write procedures as skills.

   Your refund policy, quote steps or onboarding checklist each become a skill the agent loads when it needs it.

5. 05

   5. Run the tests after every change.

   Anthropic describes a support agent where folding work into one loop broke a process step, and it showed up in only one run out of two. Run each test more than once.

Read with care

This guidance comes from Anthropic's work with enterprise customers. It is advice, not a benchmark you can repeat at home, and results depend on the model, the task and your data. Anthropic also published a migration kit on GitHub that it describes as a reference, not a maintained product. And "fewer agents" does not mean "fewer controls": approvals, logs and access limits still matter as much as ever.

This week

1. 01

   1.

   List every agent you have running or planned, and what each one does.

2. 02

   2.

   Pick one job and write 10 test tasks with the right answers.

3. 03

   3.

   Check where approvals live. If they are only in the prompt, move them into the tools.

Our take

Start with one agent you can test and explain. Add a second when a test proves you need it, not because the diagram looks more impressive. That keeps the build cheaper, faster and easier to govern.

Sources

Building agents: when to fold subagents into your main agent loop (Anthropic Applied AI team guidance; the page shows no date, X posts sharing it are dated October 22, 2026; checked October 26, 2026)

single-agent-loop-kit (Anthropic, reference migration kit linked from the guide; checked October 26, 2026)