Multi-Agent System

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.

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.