Multi-agent systems explained: one agent or several?
One AI agent or a team of specialists? How multi-agent systems split work, hand off and share memory, when they beat a single agent and when they don't.
Use one AI agent for simple or strictly step-by-step jobs, and several specialist agents when the work splits into parts that need different tools, different permissions, or can run at the same time. A multi-agent system is a group of agents, each with its own role, instructions and tools, coordinated by a lead agent or by hand-offs, usually with some shared memory. The evidence cuts both ways: Anthropic's multi-agent research system beat a single agent by 90.2% on its internal research test but used about 15 times the tokens of a chat, while Google researchers found that every multi-agent setup they tested made strictly sequential tasks 39% to 70% worse.
Below: what a multi-agent system is, the main ways agents work together, the trade-offs, a decision table, and how agent teams run in KutrimBots.
Should I use one AI agent for everything or several specialists?
Start with one, and split only when you hit a real limit: too many tools and rules for one agent to follow, data one agent shouldn't see, or work that can run in parallel.
What the research says, in numbers:
- Several agents can win on broad work. Anthropic's engineering write-up (June 2025) reports that a lead agent with parallel subagents outperformed a single agent by 90.2% on its research evaluation, and that parallel work cut research time by up to 90% on complex queries.
- It costs more. In the same post, single agents used about 4 times the tokens of a chat, and multi-agent systems about 15 times. Token usage alone explained 80% of the difference in performance on one browsing benchmark.
- Order-dependent work suffers. A Google Research study (January 2026) of 180 agent configurations found multi-agent coordination improved a parallel financial-analysis task by 80.9%, but degraded a step-by-step planning task by 39% to 70%. Agents working independently amplified errors 17.2 times; a central coordinator kept that to 4.4 times.
What is a multi-agent system?
It is 2 or more AI agents, each a language model working in a loop with its own instructions, tools and memory, coordinated towards one goal.
A single agent already plans, uses tools, remembers and asks for approval; our guide to how AI agents work explains those parts. A multi-agent system adds 3 things on top:
- Roles: each agent has a narrow job, such as research, sales or finance.
- Coordination: a rule for who does what next, either a lead agent assigning work or agents handing off to each other.
- Shared context: a common memory or thread so agents don't contradict each other.
Multi-agent systems vs agentic AI: "agentic AI" is the broad idea of AI that takes actions towards goals. A multi-agent system is one way to build it, using several agents instead of one.
Examples in real life: Anthropic's research feature (a lead agent plus searching subagents plus a citation checker), coding tools that send a helper agent to search a codebase, and small business teams where a research agent finds leads and a sales agent follows them up.
How do multiple AI agents work together?
Through a coordination pattern, plus a way to share what they know. Most systems use one of 4 patterns:
| Pattern | How it works | Good for | Watch out for |
|---|---|---|---|
| Lead and workers (orchestrator) | A lead agent splits the job, sends parts to worker agents, then combines results | Research across many sources; many independent checks | Cost; vague briefs make workers duplicate effort |
| Hand-off (pipeline) | Agent A finishes its part and passes the work to agent B | Lead → qualification → quote; draft → review | Context lost at each hand-off |
| Independent in parallel | Agents work alone; results are merged at the end | Bulk tasks with no dependencies | Errors multiply with nobody checking |
| Shared channel | Agents and people work in one thread and call each other by name | Small teams that want to see every step | Needs clear ownership of each task |
Shared memory is the other half. Agents can keep a common store of facts so the sales agent and the support agent quote the same price, and pass short references to saved work instead of copying long outputs. In Anthropic's system, the lead agent saves its plan to memory because anything beyond 200,000 tokens of context gets cut off, and subagents write results to storage and hand the lead a pointer.
What are the advantages of multi-agent systems?
The main gains are focus, safer permissions and speed on work that can be split.
- Focus. A short, narrow role card is easier for a model to follow than one giant prompt covering sales, HR and finance.
- Separate permissions. The HR agent can see salaries; the agent answering your website visitors never needs to.
- Parallel work. Independent lookups run at the same time instead of one after another.
- Right model for each job. Anthropic's system used Claude Opus 4 as the lead and the lighter Claude Sonnet 4 for subagents.
- Clear accountability. Each agent's log shows exactly what it did and why.
What are the challenges of multi-agent systems?
Cost, lost context and coordination failures, in roughly that order.
- Cost: about 15 times the tokens of a chat in Anthropic's system, as above.
- Lost context: Cognition, which builds the Devin coding agent, argues in "Don't Build Multi-Agents" (June 2025) that agents should share full context, because every action carries hidden decisions. Its example: two sub-agents asked for parts of a Flappy Bird clone return a Mario-style background and a bird that doesn't match.
- Coordination failures: a UC Berkeley-led study, "Why Do Multi-Agent LLM Systems Fail?" (NeurIPS 2025), sorted failures from 1,600+ annotated runs across 7 frameworks into 14 failure modes in 3 groups: unclear specifications, agents misaligned with each other, and weak checking of results. Two annotators agreed strongly on the labels (Cohen's kappa 0.88), and the authors trace many failures to system design, not just model limits.
- Sequential work gets worse: see the Google results above.
- Harder debugging: with 5 agents, finding which one went wrong takes longer.
When is one agent better, and when are several?
Use this table as a rule of thumb:
| Situation | One agent | Several agents |
|---|---|---|
| A personal assistant for your own inbox and calendar | Better | |
| One piece of work that must be done in order, such as a report or a code change | Better | |
| A tight budget and a light workload | Better | |
| Separate departments with separate data, such as sales and HR | Better | |
| Customers talk to an agent that must never see internal finance | Better | |
| Broad research across many sources at once | Better | |
| Jobs that run on their own schedules: morning briefs, follow-ups, month-end | Better |
When unsure, start with one agent and split it the first time its instructions get long enough to contradict themselves.
How do multi-agent teams work in KutrimBots?
As a small staffed office: agents with written roles, channels where they hand work to each other, one shared memory, and your approval on anything risky.
- Role cards. Each agent has a title, a mission, what's in scope and what's out. The Research analyst, for example, has "sending email or messages" and "spending money" out of scope. Tools a role doesn't need are switched off, and since every tool left on is described to the model at each step, narrow agents also cost less per step.
- Channels with @mentions. You put agents in a channel for a workflow. You @mention who should act, or let KutrimBots pick, and every step is visible.
- Hand-offs that sort themselves out. When an agent is asked for something outside its role, it forwards the task to a teammate in the channel who has the right tool; if nobody does, it asks you to allow the tool for this task or always; or it proposes hiring a new teammate, which you approve on a card.
- Super Brain as shared memory. Company facts and customer notes live in separate spaces, the newest value of a fact wins, and each agent's personal memory stays private until you allow another agent to read it. Super Brain has the details.
- Approvals per action. By default, anything that pays, posts publicly, deletes or contacts someone new waits for you, whichever agent asks.
- Team size by plan, as of November 2026:
| Plan | Price a month | Agents | AI usage included | Active automations |
|---|---|---|---|---|
| Free 7-day trial | $0 today | 3 | $2 | 5 |
| Solo | $20 | 5 | $10 | 20 |
| Team | $49 | 20 | $25 | 80 |
| Business | $89 | 50 | $50 | 200 |
How to set up a small multi-agent team:
- Start with one agent for the job that eats most of your time; how to set up your first AI team walks through it.
- Write its role card in plain words: mission, in scope, out of scope.
- Add a second agent only at a real limit: different tools, different permissions or parallel work.
- Create a channel for the workflow and add both agents.
- Put shared facts in Super Brain once: prices, policies, who your customers are.
- Test one hand-off with an @mention and read the log.
- Review weekly. Merge agents that rarely act; split one whose instructions keep growing.
Example scenario
Example scenario: a textile exporter in Surat on the Solo plan with 3 agents: Atlas (research), Mira (sales) and Kabir (finance).
- Monday, 9:00 AM: in the #new-buyers channel, the owner writes: "@Atlas find 10 garment importers in the UAE that buy cotton fabric." Atlas returns 10 companies with websites and buyer roles in about 25 minutes and saves a note to Super Brain.
- Monday, 9:30 AM: Atlas @mentions Mira with the list. Mira drafts 10 first emails from the price list in Super Brain. Because the contacts are new, all 10 wait for approval; the owner approves 8 and edits 2 from his phone.
- Thursday: a buyer asks for 2,000 metres at ₹185 a metre. Mira hands the quote to Kabir, who prepares a ₹3,70,000 quote (plus GST) in Invoices & Billing for the owner to check.
- Friday: Atlas is asked to "send the quote". Sending is out of its scope, so it forwards the task to Mira instead of trying.
- Usage: three agents take more steps than one would. Solo's $10 of included AI usage covers roughly 3,300 average agent replies on the included models, by KutrimBots' own estimate, and a $10 top-up carries over. What an AI agent costs compares pricing models.
What can't a multi-agent team do?
It can't fix a badly defined job, and it adds overhead you should only pay for when the work needs it.
- More agents isn't more intelligence. Each hand-off can drop context; long, order-dependent work is often better with one owner.
- Costs rise with agents and steps. Watch usage for the first month.
- Approvals can pile up. 3 agents asking permission means more cards; allow trusted contacts and tools "always" once you're confident.
- Unclear roles cause loops. If two agents both think a task is theirs, or neither does, fix the role cards.
KutrimBots has a free 7-day trial on Solo: add a card or UPI, nothing is charged today, and you can run up to 3 agents before it becomes Solo at $20 a month (as of November 2026). See plans and team sizes.
Frequently asked questions
What is an example of a multi-agent system in real life?
Anthropic's research feature uses a lead agent that sends subagents to search in parallel and a separate agent to check citations. In a small business, a research agent finding leads and handing them to a sales agent is a simple multi-agent system.
Do multi-agent systems cost more than a single agent?
Usually, yes. Anthropic reports its multi-agent system used about 15 times the tokens of a chat, against about 4 times for a single agent, so use several agents only where the extra quality or speed pays for itself.
What is AI agent orchestration?
Orchestration is how work is divided and passed between agents: a lead agent assigning tasks to workers, agents handing off in sequence, or agents working in a shared channel. It also covers what context and memory each agent gets.
Multi-agent systems vs agentic AI: what's the difference?
Agentic AI is the general idea of AI that acts towards goals with tools. A multi-agent system is one design for it, where several specialised agents coordinate instead of one agent doing everything.
How many AI agents should a small business start with?
One. Add a second when the first agent's role gets too broad, needs data it shouldn't see, or has work that could run in parallel. On KutrimBots Solo you can grow to 5 agents without changing plans.