# Multi-Agent System

AI Glossary · Agents & Automation
Canonical URL: https://aisat.uk/glossary/multi-agent-system
Author: Lucy Gilmore, Assistant General Counsel, Product, Privacy & AI at Sprinklr | Practical Privacy, Product & AI Governance & Defensible Risk Decisions (https://www.linkedin.com/in/lucy-gilmore/)
Breadcrumb: [Glossary](https://aisat.uk/glossary) > [Agents](https://aisat.uk/glossary?category=agents) > Multi-Agent System

## TL;DR

> **A multi-agent system** (MAS) is an architecture in which multiple specialised AI agents with distinct instructions, tools, and context windows collaborate to solve a complex problem. By dividing responsibilities across orchestrator and worker agents, it prevents context overload on large tasks.

## How Multi-Agent System works in practice

When a single AI agent is given fifty tools, a massive system prompt, and a multi-stage goal, its accuracy drops sharply: tool selection becomes unreliable and intermediate search dumps fill the context window with noise. A multi-agent system applies the software engineering principle of separation of concerns, partitioning the problem across smaller, domain-focused agents.

Two coordination topologies dominate production systems. In an orchestrator-worker (or supervisor) pattern, a lead planner agent decomposes the user's objective into discrete subtasks, dispatches them in parallel or sequence to specialist sub-agents (each running in its own isolated context window with only the 3 to 5 tools it needs), and synthesises their compact return summaries. In a handoff (or peer-to-peer graph) pattern, control transfers explicitly from a triage agent to a specialist agent via structured state updates or over the A2A protocol.

For example, a London investment research platform runs a multi-agent due-diligence pipeline: when an analyst requests a briefing on a FTSE 250 acquisition target, a supervisor agent dispatches parallel sub-agents to parse Companies House filings, earnings call transcripts, and CMA regulatory notices, before a verifier agent cross-checks every financial ratio against the source tables.

## Why it matters and trade-offs

Context isolation is the real superpower of a multi-agent system. If a sub-agent reads 80,000 tokens of raw log files or legal annexures to find three relevant clauses and returns only a 400-token structured summary to the supervisor, the main reasoning loop stays clean, fast, and focused.

However, multi-agent systems are frequently over-engineered for tasks that a single agent or a linear agentic workflow could handle in half the time. Every inter-agent handoff adds network latency, duplicates system-prompt token overhead, and creates 'telephone game' information loss if structured schemas are not enforced between agents. Always start with the simplest single-agent workflow and split out sub-agents only when context pollution or tool-routing errors show up in your evals.

## Related AI concepts

- [AI Agent](https://aisat.uk/glossary/ai-agent): An AI Agent is a software system powered by a large language model that perceives environment context, maintains state, formulates multi-step plans and executes actions across external tools, APIs and databases to complete a delegated goal.
- [A2A (Agent2Agent Protocol)](https://aisat.uk/glossary/a2a-protocol): A2A (Agent2Agent Protocol) is an open interoperability standard launched by Google in April 2025 and donated to the Linux Foundation in June 2025 that enables AI agents across different vendors and frameworks to discover capabilities, delegate tasks and exchange artefacts.
- [Agentic Workflow](https://aisat.uk/glossary/agentic-workflow): An Agentic Workflow is a multi-step operational or engineering process in which a language model dynamically plans steps, calls external tools, critiques its own output and branches based on intermediate results within an orchestrated pipeline.
- [Context Engineering](https://aisat.uk/glossary/context-engineering): Context Engineering is the engineering discipline of dynamically selecting, structuring, compressing and ordering the system instructions, retrieved knowledge, tool definitions, memory and examples placed inside a language model's context window at inference time.