On this pageJump to a section
Two words get used as if they mean the same thing: workflow and agent.
They do not, and the difference decides how much you will spend, how much you can trust the result and how much of your Friday you lose to debugging.
Anthropic draws the line cleanly. Workflows are systems where models and tools are orchestrated through predefined code paths. Agents are systems where models dynamically direct their own processes and tool usage.2 I covered the same split in Agents vs Workflows. On June 13, 2025, the company published a detailed account of how it built a multi-agent research system.1 Read it with a small business in mind, and it is a masterclass in when to use one and when to leave it alone.
How does Anthropic's multi-agent research system work?
Anthropic describes an orchestrator-worker pattern. A lead agent takes the user's question, plans an approach and spawns several subagents. Those subagents run in parallel, each with its own context window, tools and instructions, searching iteratively and returning what they find to the lead.1 For the plumbing that lets agents talk to each other, see Agent2Agent and MCP.
- QuestionThe user asks something broad
- Lead agentPlans the approach and spawns helpers
- SubagentsWork in parallel, each on one slice
- Citation agentAttributes claims to sources
- ReviewA person reads before anything goes out
The cast
- Lead agent. Plans the work, delegates and synthesizes the final answer. It saves its plan to memory because its context can run past what it can hold.1
- Subagents. Specialists that each take one slice of the question.
- Citation agent. After the research is done, a separate agent processes the documents and the report to identify where claims should be attributed to sources.1
What it achieved
In Anthropic's internal research evaluation, a system with Claude Opus 4 as the lead and Claude Sonnet 4 subagents outperformed a single-agent Claude Opus 4 by 90.2%.1
90.2%
That is Anthropic's own evaluation, on its own task, so read it as evidence the pattern can help on broad research, not as a promise for your business.
What does it cost?
This is the part most excited blog posts skip. Anthropic reported that agents typically use about 4 times the tokens of chat interactions, and multi-agent systems use about 15 times as many.1
The company also found that on one benchmark, token usage alone explained 80% of the variance in performance.1 In plain English: much of the improvement came from doing more thinking and searching. More work costs more money. A task has to be valuable enough to justify the bill.
What can a small business learn from the design?
Anthropic shared eight prompt engineering principles. Five translate straight into automations an operator could build.
1. Teach delegation
Subagents did poorly with vague instructions. The fix was giving each one an objective, an output format, guidance on tools and clear task boundaries, so they did not duplicate each other's work.1 Any time you hand work to an AI helper, or a human, those four things belong in the brief.
- Objective. What exactly should this helper find out or produce?
- Output format. What should the answer look like, so it can be combined with others?
- Tools and sources. Where should it look, and where should it not?
- Boundaries. What is out of scope, so two helpers do not do the same job?
2. Scale effort to complexity
Anthropic embedded rules for how much effort a question deserves: a simple fact needs one agent and a few tool calls, while complex research warrants many subagents.1 Do not build a team to answer a one-line question.
3. Design tools carefully
Each tool needs a distinct purpose and a clear description, because bad descriptions send agents down the wrong path.1 If you connect an assistant to your CRM, calendar and documents (see Claude Integrations), name and describe what each one is for.
Erik Schluntz, a researcher at Anthropic, made the same point in a video posted on February 13, 2025. People write beautiful prompts, he said, then hand the model bare-bones tools, and "you need to be prompt engineering in the descriptions of your tools themselves."3
4. Start wide, then narrow
Agents did better beginning with short, broad queries and focusing progressively.1 That works for people too.
5. Parallelize what can run in parallel
Anthropic reported that running subagents and tool calls in parallel cut research time by up to 90% for complex queries.1 The upside is real when the pieces are independent.
How do you evaluate an AI system without a lab?
You do not need a research team. Anthropic's advice is practical: start small. About 20 queries drawn from real use were enough in early development, since large changes show up clearly in small samples.1
Judge outcomes, not paths
Agents can take different valid routes to the same answer. Anthropic recommends judging whether the result is right and the process reasonable, rather than checking every step.1 Write a short rubric for your own automation: Is it accurate? Does it cite sources? Is it complete? Did it use the right information?
Keep a human in the test
Anthropic noted that human testing found edge cases automation missed, such as the system favoring low-quality content farms over more authoritative sources.1 For anything customer-facing, a person should read the output before you trust it.
When should you not use multiple agents?
Anthropic is direct about this. Multi-agent systems are a poor fit for:1
| Poor fit | Better choice |
|---|---|
| Tasks where all agents need the same shared context | A single agent with the full picture |
| Work with many dependencies between steps | A fixed workflow that runs step by step |
| Tasks with few parts that can run in parallel | A single call or a short chain |
| Low value work, given the token cost | A cheaper, simpler automation |
In its earlier guide to building agents, Anthropic recommends finding the simplest solution possible and only increasing complexity when needed, noting that agentic systems often trade latency and cost for better task performance and that many tasks need only a single optimized call.2
finding the simplest solution possible, and only increasing complexity when needed
What would a small business version look like?
A weekly market brief is a good candidate: broad, parallel and valuable, and the audit habits in Deep Research for Operators apply. A lead agent splits the question into parts, such as competitors, pricing, news and customer questions. Helpers research each part, a final pass checks citations and a person reads the result before it goes anywhere.
A reminder email sequence is not. It is a fixed workflow with a few steps, and it should stay one.
With no team to delegate to, I care about what AI can take off my plate. My rule is boring and it works: start with the simplest version, prove it on real cases, then add complexity only where the result is worth the cost. Use agents for judgment, workflows for routine.
The goal is not more agents. It is fewer things that need your attention.
Deciding what to automate and what to leave alone is a question I work through with clients.
Keep building.Ekai
Questions people ask
What is a multi-agent AI system?
It is a setup where a lead AI agent plans a task, hands pieces of it to several helper agents that work in parallel, then combines their results. Anthropic calls its version an orchestrator-worker pattern.
What is the difference between a workflow and an agent?
Anthropic defines workflows as systems where models and tools are orchestrated through predefined code paths, and agents as systems where models dynamically direct their own processes and tool usage.
Are multi-agent systems more expensive?
Yes. Anthropic reported that agents use about 4 times the tokens of chat interactions and multi-agent systems about 15 times, so the task has to be valuable enough to justify the cost.
When should a small business not use multiple agents?
When the task needs shared context, has many dependencies between steps, has few parallel parts or is low value. In those cases a single agent or a fixed workflow is usually better.
Sources
- 1
- 2
- 3
