Skip to content
Blentera

Resources, AI agents

Multi-agent systems: patterns, trade-offs and governance

A multi-agent system splits work between several AI agents that pass tasks and results to each other. This guide covers the common patterns, when they beat a single agent, what they cost and how to keep them under control.

Updated 20 November 2026. 9 minute read.

In short

  • A multi-agent system is a group of AI agents, each with its own instructions, tools and context, coordinated by another agent or by code to complete a task together.
  • The common patterns are a manager agent with workers, hierarchical teams, handoffs between specialists, and review of one agent's work by a second agent.
  • Start with one agent. Add agents when work splits into independent parts, needs more context than one agent handles well, or needs separate permissions. Keep a single agent when steps depend tightly on each other or several agents would change the same thing.
  • More agents cost more. Anthropic measured multi-agent systems using about fifteen times the tokens of a chat, and errors can compound between agents.
  • Governance must follow the action across agents: a role and an owner per agent, approval rules per tool, budgets per agent and one log that traces each task end to end.

What a multi-agent system is

A multi-agent system is a set of AI agents that each handle part of a task and pass work between them. Each agent has its own instructions, tools and context. One agent may research, another may write and a third may check the result.

Coordination happens in one of two ways. The OpenAI Agents SDK documentation describes them clearly. Either a language model decides which agent acts next, or code decides. Model-led coordination handles open-ended tasks. Code-led coordination is more predictable in speed, cost and performance. Most production systems mix the two.

Agents built on different platforms can also work together through open protocols. The Agent2Agent (A2A) protocol, created by Google and contributed to the Linux Foundation in June 2025, is designed for communication between agents. The Model Context Protocol (MCP) connects an agent to tools and data. The two address different layers.

Common patterns

A manager agent with workers

A manager agent, also called an orchestrator, breaks the task into subtasks, delegates them to worker agents and combines the results. Anthropic calls this orchestrator-workers. Its research feature uses it: a lead agent plans the research, starts subagents that search in parallel and then synthesises what they return. The OpenAI Agents SDK calls the same idea agents as tools. The manager keeps control and calls specialist agents as if they were tools. Use this pattern when one agent should own the final answer.

Hierarchical teams

A hierarchy adds layers. A top-level agent delegates to team leads, which delegate to their own workers. This mirrors an org chart and keeps each agent's context small. Anthropic describes how subagents can work in clean context windows and return a condensed summary to the lead. Each extra layer adds model calls, waiting time and another point where details can be lost, so keep hierarchies shallow.

Handoffs

In a handoff, one agent passes the task to another, which takes over. A triage agent might pass a refund request to a billing specialist. In the OpenAI Agents SDK, a handoff appears to the model as a tool, and the specialist becomes the active agent for the rest of the turn. OpenAI's guide to building agents calls this the decentralised pattern, in which agents work as peers. Handoffs suit routing, where the specialist should respond directly.

Review by a second agent

One agent produces the work and another checks it against set criteria, sending feedback until it passes. Anthropic calls this evaluator-optimiser and recommends it when the criteria are clear and another round adds measurable value. Cognition, which builds coding agents, reported in 2026 that its review agent worked best when it did not share the coding agent's context beforehand, so it looked at the work afresh. A reviewer agent is a quality check. It does not replace employee approval for risky actions.

When several agents help, and when one is better

Start with one agent. OpenAI's guide recommends maximising a single agent's capabilities first, and splitting only when it fails to follow complicated instructions or keeps choosing the wrong tools. Anthropic gives similar advice: use the simplest system that works, and add complexity only when it clearly improves results.

When several agents help

  • The work splits into independent parts that can run in parallel, such as researching several companies at once. On Anthropic's internal research evaluation, its multi-agent system outperformed a single Claude Opus 4 agent by 90.2%.
  • The task needs more information than one agent handles well. Each subagent explores with a clean context and returns a short summary.
  • Different steps need different tools, permissions or models. Separate agents keep each one's access narrow.
  • An independent check adds value, as with a reviewer agent.

When one agent is better

  • Steps depend tightly on each other, or every agent would need the same context. Anthropic says such tasks are not a good fit for multi-agent systems today, and notes that most coding tasks have fewer truly parallel parts than research.
  • Several agents would change the same document, code or record. Cognition argued in 2025 that actions carry implicit decisions, so parallel agents that cannot see each other's decisions produce conflicting work. In 2026 it reported the best results when changes stay with one agent and the others contribute analysis.
  • The task is small or speed matters. Every extra agent adds model calls, waiting time and cost.
  • You cannot yet measure the single agent's quality. Without a baseline, you cannot tell whether more agents help.

Cost and reliability trade-offs

Multi-agent systems spend more to achieve more. Anthropic measured that agents use about four times the tokens of a chat interaction, and multi-agent systems about fifteen times. In its analysis, token usage alone explained 80% of the variance in performance on the BrowseComp benchmark. Better results often come from spending more, so the task must be worth the extra cost.

Reliability fails in recognisable ways. A UC Berkeley study presented at NeurIPS 2025 analysed traces from popular multi-agent frameworks and grouped 14 failure modes into three categories:

  • specification and design problems, such as an agent ignoring its role or the task's requirements, repeating steps or not knowing when to stop;
  • misalignment between agents, such as withholding information, ignoring another agent's input or drifting from the task;
  • weak verification, such as stopping too early or checking the result poorly or not at all.

Errors also compound. Anthropic notes that agents keep state, so one failed step can send a run down a different path. Its early versions started 50 subagents for simple queries until it wrote scaling rules into the prompts. Full production tracing was what let it diagnose failures systematically.

  • Set a budget per run and per agent, and a limit on steps and retries.
  • Tell the manager how many workers to start for each kind of task, and when to stop.
  • Pass structured results between agents, with sources, instead of long free text.
  • Compare the whole system with a single agent on the same evaluation tasks before you adopt it.

Governance: roles, approvals and logging across agents

Governance gets harder as agents multiply, because one action can pass through several agents before it reaches an app. The OWASP Top 10 for agentic applications, published in December 2025, lists identity and privilege abuse and rogue agents among the main risks. It recommends a separate identity per agent, short-lived and narrowly scoped credentials, behaviour monitoring and a way to stop an agent at once.

Roles

Give every agent a defined role, a named employee as owner and a place in the org chart: what it does, which agent it reports to and which agents it may delegate to. A worker should not inherit all of its manager's access by default. Each agent gets only the tools its role needs.

Approvals

Approval rules must follow the action, not the agent that started the task. If a manager agent asks a worker to send an email, the rule for sending email applies to the worker. Set rules per agent, project, app or tool, so that a sensitive tool needs approval whichever agent calls it. A reviewer agent can catch errors before an employee sees the work, but it does not replace employee approval for irreversible or external actions.

Budgets

Set budgets per agent and per project, so a manager agent cannot run up costs through its workers. Warn before the limit and pause at it.

Logging

Log each step with the agent that took it, the agent or employee that asked for it, the tool used, any approval and the cost. Link all steps of one task with a shared reference, so you can follow the task across agents. For high-risk AI systems, Article 26(6) of the EU AI Act requires deployers to keep the logs under their control for at least six months, unless other law provides otherwise.

A hypothetical example: a research team of agents

The following example is hypothetical. A consultancy wants weekly briefings on five competitors. A single agent struggled to keep all five in its context, so the company tests a small team.

  1. A manager agent receives the request, splits it into one subtask per competitor and sets a budget for the run.
  2. Five worker agents research one competitor each, in parallel, with read-only search tools. Each returns a structured summary with sources.
  3. The manager writes the briefing. It is the only agent that writes, so there are no conflicting edits.
  4. A reviewer agent checks every claim against the cited sources, without seeing the manager's reasoning, and returns a list of problems.
  5. An employee approves the final briefing before it is sent to the client, because it leaves the company.
  6. The log links every step to the run, so the owner can see which worker found what and what each step cost.

Before adopting the team, the company runs the single agent and the team on the same set of past briefing requests. It compares quality, cost per briefing and time, and keeps the team only if quality rises enough to justify the cost.

How Blentera helps

Blentera is a platform where companies run AI agents with employees in charge. Each agent has a role in the org chart, so it is clear what each agent does and who is responsible. Approval rules can be set per agent, project, app or tool. Budgets warn at 80% and pause at the limit. Agents share one company memory, which stays when you switch AI models, and an activity log records what each agent does. Native agents work alongside OpenClaw, Hermes and Cursor Cloud agents, with connectors to apps and MCP servers. Blentera is hosted in the EU.

Questions

Sources

  1. How we built our multi-agent research system, Anthropic Engineering
  2. Building effective agents, Anthropic Engineering
  3. Effective context engineering for AI agents, Anthropic Engineering
  4. Agent orchestration, OpenAI Agents SDK documentation
  5. Handoffs, OpenAI Agents SDK documentation
  6. A practical guide to building agents, OpenAI
  7. Don't build multi-agents, Cognition, June 2025
  8. Multi-agents: what's actually working, Cognition, 2026
  9. Why do multi-agent LLM systems fail? Cemri et al., NeurIPS 2025, arXiv:2503.13657
  10. Google Cloud donates A2A to Linux Foundation, Google for Developers
  11. OWASP Top 10 for Agentic Applications for 2026, OWASP GenAI Security Project
  12. Regulation (EU) 2024/1689 (AI Act), consolidated text of 27 July 2026, EUR-Lex

This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.

Get started

Put your first agent to work in minutes.

Three steps, and your first agent is working for your company.

  1. Step 1

    Create an account

    Sign up with your work email and invite your team. No sales call needed.

  2. Step 2

    Add your first agent

    Create an agent, give it a role, and connect your own AI key and the apps it may use.

  3. Step 3

    Give it a task

    Ask in Blentera or your chat app, and follow the work until it is done.