A multi-agent system uses multiple agents with different responsibilities to work on one larger goal.

Instead of one agent doing everything, work is divided among specialized agents.

Simple example

Goal:

Review a backend pull request.

Possible agents:

• Code agent: checks logic and maintainability
• Security agent: looks for authorization and injection problems
• Test agent: finds missing tests and edge cases
• Coordinator: combines the findings

flowchart TB
    A["Coordinator"] --> B["Code review agent"]
    A --> C["Security agent"]
    A --> D["Test agent"]
    B --> E["Combined result"]
    C --> E
    D --> E

Why use multiple agents?

It can help when:

• work can be clearly divided
• different tasks need different instructions or tools
• independent parts can run in parallel
• one agent should review another agent's result

Common coordination patterns

Coordinator and specialists

One coordinator divides the task, sends work to specialist agents, and combines the results.

Parallel workers

Several agents perform independent tasks at the same time.

Example:

Research API options
Research database options
Research security requirements

Reviewer pattern

One agent creates a result, and another checks it against clear criteria.

Shared state and handoffs

Agents need a structured way to communicate:

{
  "task": "Check authorization in order routes",
  "findings": [
    {
      "file": "orders.js",
      "issue": "Ownership is not checked",
      "severity": "high"
    }
  ]
}

Structured handoffs are more reliable than sending a long, unorganized chat transcript.

The coordinator should track:

• task assigned to each agent
• completion status
• outputs and evidence
• errors
• total cost and time

More agents do not automatically mean better results

Multi-agent systems add:

• more model calls
• higher token cost
• longer execution time
• duplicated work
• conflicting answers
• harder debugging
• more security boundaries

For many tasks, one agent with good tools is simpler and better.

When not to use multiple agents

Avoid them when:

• the task is small
• steps depend heavily on one another
• one agent already performs reliably
• coordination costs more than the work
• different agents would use the same prompt and tools

Start with the simplest design. Add agents only when specialization or parallel work produces measurable improvement.

Security rules

Each agent should receive only:

• the tools it needs
• the data it is allowed to access
• the task it must perform

A report-writing agent should not receive a payment tool.

A multi-agent system is a team of specialized AI workers. Use it only when dividing the work is clearer and more valuable than using one capable agent.