What is the difference between a single agent and a multi-agent system?
A single agent is one model with one set of instructions that calls all the tools itself. A multi-agent system splits the job across several specialised agents, coordinated by a manager agent, by handoffs between them, or by a fixed workflow.
Microsoft's architecture guidance puts it this way: single-agent systems "consolidate all logic into a single agent", which "simplifies implementation, reduces operational overhead, and offers a more predictable execution model", while multi-agent systems "divide responsibilities across multiple specialized agents", which brings modularity "but requires additional coordination and orchestration" (Microsoft Cloud Adoption Framework, single-agent or multi-agent, updated February 27, 2026, checked October 30, 2026). Choosing between the two is one of the first design decisions in our AI agent development work, and it changes the build time, the running cost and how easy the agent is to test.
One point of vocabulary matters. Anthropic separates workflows, "systems where LLMs and tools are orchestrated through predefined code paths", from agents, "systems where LLMs dynamically direct their own processes and tool usage" (Anthropic, Building effective agents, checked October 30, 2026). Many systems sold as "multi-agent" are a workflow that calls several agents in a fixed order. That is often the right design, and it is a different thing from agents deciding among themselves who works next. If your real question is whether you need an agent at all, our AI workflow automation page compares agents with plain workflows.
Should you start with a single agent or a multi-agent system?
Start with one agent in almost every case. OpenAI, Anthropic and Microsoft all recommend getting the most out of a single agent first and adding agents only when testing shows a limit that better tools, prompts or retrieval cannot fix.
Where to start, per the provider guides
The three providers say it in their own words:
- OpenAI: "Our general recommendation is to maximize a single agent's capabilities first. More agents can provide intuitive separation of concepts, but can introduce additional complexity and overhead, so often a single agent with tools is sufficient." (OpenAI, A practical guide to building agents, checked October 30, 2026)
- Anthropic: "We recommend finding the simplest solution possible, and only increasing complexity when needed."
- Microsoft: transition to multi-agent "only when testing reveals limitations that cannot be resolved through single-agent optimization". Its guide adds that multi-agent designs "are often chosen based on untested assumptions about complexity or performance."
Microsoft names one exception: start with several agents when the use case crosses security and compliance boundaries, involves several teams, or has planned growth across many functions. Everything else starts with a single-agent test.
When does a multi-agent system make sense?
When the job crosses a security or compliance boundary, when one agent keeps picking the wrong tool among many similar ones, when its instructions turn into a maze of conditions, or when the work splits into independent parts that can run at the same time.
We collected the triggers each provider names into one checklist. If none of these is true for your job, one agent is the better build.
| Signal | What the provider says | Source |
|---|---|---|
| Data must be kept apart for security or compliance | "Build multiple agents when regulations or policies mandate strict data isolation." Example: one agent prepares a transaction, another validates it. | Microsoft |
| Separate teams own separate knowledge areas | Distinct teams can maintain domain-specific agents and deploy updates independently. | Microsoft |
| The roadmap spans many functions | "Solutions spanning more than three to five distinct functions benefit from this architecture." | Microsoft |
| Instructions full of if-then-else branches | When prompts contain many conditional statements and templates get hard to scale, divide each logical segment across separate agents. | OpenAI |
| The agent picks the wrong tool | "Some implementations successfully manage more than 15 well-defined, distinct tools while others struggle with fewer than 10 overlapping tools." Split only if clearer tool names and descriptions do not help. | OpenAI |
| Broad research that splits into independent parts | Anthropic's multi-agent research system "outperformed single-agent Claude Opus 4 by 90.2%" on its internal research eval. | Anthropic |
The research result comes from Anthropic's post on how it built its multi-agent research system (June 13, 2025, checked October 30, 2026). Read it with its context: the task was open-ended research, where sub-agents search different sources in parallel. Anthropic also found that "token usage by itself explains 80% of the variance" in its BrowseComp evaluation, so much of the gain came from spending more tokens on the problem.
When is a multi-agent system the wrong choice?
When every step needs the same context, when steps depend on each other in sequence, or when speed and cost matter more than breadth. Coordination adds latency at each handoff, extra tokens, and more places for a decision to go wrong.
Anthropic names the poor fits in the same post that reports the 90.2% gain: domains that need "all agents to share the same context", tasks with "many dependencies between agents", and most coding work, which involves "fewer truly parallelizable tasks than research". Walden Yan of Cognition went further in June 2025. Its two principles are "share context, and share full agent traces, not just individual messages" and "actions carry implicit decisions, and conflicting decisions carry bad results", and it concluded that running multiple agents in collaboration "only results in fragile systems" (Cognition, Don't build multi-agents, checked October 30, 2026).
Microsoft lists the operating costs of a multi-agent design plainly: each interaction needs protocol design, error handling and state synchronisation; each agent needs its own prompts, monitoring and debugging; security surfaces grow with more credentials and data transit points; costs rise because "each agent processes redundant context"; and "latency accumulates at each handoff point".
For most business jobs we see (qualify a lead, answer from a help center, update a CRM, book a slot), the work is sequential and shares one context. One agent with good tools, or a workflow with one agent step, does it with fewer failure points. For those jobs, start with the ready-made AI agents for voice, chat and booking before designing anything custom.
How much more does a multi-agent system cost to run?
Expect several times the tokens. Anthropic measured agents at about 4 times the tokens of a chat and multi-agent systems at about 15 times, which suggests a multi-agent version of a research job can cost roughly 3.75 times its single-agent version.
Token use compared with a chat
Anthropic's figures: "agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats." Dividing 15 by 4 gives 3.75, the rough multiplier from one agent to a multi-agent system on that kind of work. We applied it to the research-and-report workload priced on our AI agent development page (10 model calls per task, 100 million input and 4 million output tokens per 1,000 tasks), using the Claude pricing page read October 30, 2026.
| Model (in / out per 1M tokens) | Single agent, per 1,000 tasks | Multi-agent estimate (x 3.75) | Extra per 1,000 tasks |
|---|---|---|---|
| Claude Haiku 5.5 ($0.10 / $0.50) | 100 x $0.10 + 4 x $0.50 = $12.00 | $45.00 | $33.00 |
| Claude Sonnet 5.5 ($2 / $10) | 100 x $2 + 4 x $10 = $240.00 | $900.00 | $660.00 |
| Claude Opus 5.5 ($4 / $20) | 100 x $4 + 4 x $20 = $480.00 | $1,800.00 | $1,320.00 |
Treat this as an order of magnitude, not a quote. The 4x and 15x figures are Anthropic's averages for its own research system; your multiplier depends on how many sub-agents run and how much context each one receives. A fixed workflow that calls narrow agents one after another, each with only the documents for its section, can cost far less than 3.75x because no agent carries the whole history. The only reliable number is the one you measure by running your own test set both ways.
What are the main multi-agent patterns?
A manager agent that calls specialist agents as tools, specialist agents that hand the conversation to each other, an orchestrator that splits work across parallel workers, and a fixed workflow that runs narrow agents in order. The last one is the easiest to test.
| Pattern | How it works | Good for | Watch for |
|---|---|---|---|
| Manager (agents as tools) | OpenAI: a central "manager" agent "coordinates multiple specialized agents via tool calls" | One voice to the user, specialists behind it | The manager becomes the bottleneck and the place errors hide |
| Decentralized handoffs | OpenAI: "agents can 'handoff' workflow execution to one another", passing the latest conversation state | Triage, where one agent routes to a billing, support or sales agent | Context lost between handoffs |
| Orchestrator-workers | Anthropic: a central model "dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results" | Broad research with independent parts | Token cost; workers making conflicting choices |
| Fixed workflow of narrow agents | A workflow runs each agent in a set order; each has one job and a fixed output | Reports and documents with defined sections | Less flexible when the input does not fit the sections |
The fourth pattern is the one we have shipped. In our Dutch real estate property report case study, eight AI agents run one after another in n8n. Each fills one section of the report, searches the same document store, and returns its answer through a structured output parser into its own Google Sheets tab. No agent decides who works next; the workflow does. That gave the client the benefit people want from multi-agent systems (narrow jobs, one place to look when a section is wrong) without agents negotiating with each other.
How do you decide, for your own job?
Build the single-agent version, write a test set from real cases, and measure where it fails. Add a second agent only for a failure the test set shows, and rerun the same tests to prove the split helped and was worth the extra cost.
Microsoft calls this a "comparative prototype": "execute controlled tests comparing architectures against defined success metrics". In practice:
- Write down the job. Inputs, tools, the correct result, and what the agent must never do.
- Build one agent with clearly named tools and instructions written as rules.
- Test it on 20 to 50 real cases, several trials each. Our guide on how to test an AI agent before launch covers the test set and grading.
- Read the failures. Wrong tool among similar ones? Rename and describe the tools first. Instructions contradicting each other? Split the instructions, or route to a second agent. Missing information? That is a retrieval fix, not an architecture one.
- Split only where it fails, then rerun the same set and compare pass rate, steps and cost per case.
If you already know which job you want automated but not how to structure it, an AI consulting audit does this mapping before any build starts.
More questions about single and multi-agent systems
What is an example of a multi-agent system in business?
A support desk where a triage agent reads each request and hands it to a billing agent or a technical agent, each with its own tools and permissions. Another is a report generator where several agents each fill one section from the same documents, as in our Dutch property report build. Both split the work because the parts need different tools or different data.
Is a multi-agent system more accurate than a single agent?
Sometimes, on broad research. Anthropic reports a 90.2% improvement for its multi-agent research system on its own internal eval. On sequential jobs that share one context, the providers warn that coordination adds failure points. Accuracy is a property of your job and your test set, so measure both versions on the same cases before deciding.
How many tools can one AI agent handle?
There is no fixed limit. OpenAI's guide says some implementations manage more than 15 well-defined, distinct tools while others struggle with fewer than 10 overlapping ones. The problem is similarity between tools more than their number. Clear tool names and descriptions that say when to use each tool usually fix wrong tool choices before a second agent is needed.
What is the difference between a sub-agent and a multi-agent system?
A sub-agent is an agent another agent calls like a tool and gets a result back from, which is OpenAI's manager pattern. The caller stays in charge. In a multi-agent system with handoffs, control moves to the other agent, which continues the conversation itself. Sub-agents are easier to test because one agent always owns the final answer.
Can n8n run multi-agent systems?
Yes. An n8n workflow can run several AI Agent nodes in sequence or in separate branches, each with its own model, tools and output parser. That makes the fixed-workflow pattern straightforward to build and to test. Our news piece on the n8n Agents launch covers the newer standalone agents and how they differ from the AI Agent node.
Does a multi-agent system need a different model for each agent?
No, but it is one of the real advantages. A narrow agent with a simple job can often run on a small, cheap model while only the hardest step uses a large one. The test set shows which agents pass on the smaller model. That can offset part of the extra token cost that coordination adds.
Sources and further reading
- OpenAI, A practical guide to building agents (maximize a single agent first; splitting guidelines; manager and decentralized patterns), checked October 30, 2026
- Anthropic, Building effective agents (December 19, 2024: workflows vs agents, orchestrator-workers), checked October 30, 2026
- Anthropic, How we built our multi-agent research system (June 13, 2025: 90.2% result, 4x and 15x token usage, poor fits), checked October 30, 2026
- Microsoft Cloud Adoption Framework, Choosing between a single-agent or multi-agent system (decision criteria and trade-offs, updated February 27, 2026), checked October 30, 2026
- Cognition, Don't build multi-agents (Walden Yan, June 12, 2025), checked October 30, 2026
- Claude pricing (Haiku 5.5, Sonnet 5.5, Opus 5.5 token rates), checked October 30, 2026
