TL;DR
- Multi-agent systems use multiple specialised agents coordinated by an orchestrator
- Solve three problems: context window limits, task parallelism, specialist quality
- Core patterns: orchestrator–worker, pipeline, debate/critic, and parallel fan-out
- Each agent has a defined role, tools, and scope: like hiring different specialists
- More complex to build and debug: only justified when single agent genuinely fails
Why Single Agents Fail at Complex Tasks
A single AI agent works well when the task is focused and sequential. It struggles when:
✗ Context window overflows. Processing 200 CVs, 50 competitor pages, or a full year of transaction data in one agent run fills the context window. The agent starts forgetting earlier data or refusing to process more. ✗ Tasks are independent and should run in parallel. If you need to research 20 prospects, a single agent processes them one by one (slow). Multiple parallel agents each take 5 prospects simultaneously: 4x faster. ✗ Different steps need different expertise. A sales proposal agent needs research capability, copywriting quality, and financial analysis. No single system prompt does all three well: specialist agents outperform generalist ones. ✗ Quality needs checking. A single agent reviewing its own work is unreliable. A separate critic agent reviewing the output produces measurably better results: same principle as peer review.The Four Multi-Agent Patterns
Pattern 1: Orchestrator–Worker
An orchestrator agent breaks down a goal and delegates sub-tasks to specialist worker agents. Workers complete their tasks and return results. The orchestrator synthesises everything into the final output. Implementing this pattern cleanly usually comes down to your choice of AI agent framework: LangGraph, CrewAI, or AutoGen. Each of these models orchestration and delegation differently.
Example: Sales proposal system: Orchestrator → [Research Agent, Pricing Agent, Copywriting Agent] → Orchestrator compiles proposal
Pattern 2: Pipeline
Agents process work in sequence: each agent's output becomes the next agent's input. Like an assembly line: each station does one job well.
Example: Content production: Research Agent → Outline Agent → Drafting Agent → Editing Agent → SEO Agent → Publish
Pattern 3: Parallel Fan-Out
The orchestrator spawns multiple identical agents in parallel to process independent items simultaneously. Drastically reduces processing time for large volumes.
Example: CV screening: Orchestrator splits 100 CVs into 10 batches → 10 screening agents run in parallel → Orchestrator merges ranked shortlist
Pattern 4: Critique & Revision
A generator agent produces a draft. A separate critic agent reviews it against specific criteria and provides structured feedback. The generator revises. This loop runs until quality criteria are met.
Example: Contract drafting: Drafting Agent produces clause → Legal Review Agent checks against compliance criteria → flags issues → Drafting Agent revises
Real-World Multi-Agent Systems
| Business Use Case | Pattern | Agents Involved | Build Cost |
|---|---|---|---|
| Sales proposal generation | Orchestrator–Worker | Research, Pricing, Writer, Reviewer | £20k–£40k |
| Large-scale CV screening | Parallel Fan-Out | Orchestrator + N screening agents | £12k–£22k |
| SEO content production | Pipeline | Research, Outline, Writer, Editor, SEO | £15k–£30k |
| Market intelligence monitoring | Parallel + Pipeline | Crawler agents + Analysis agent | £15k–£28k |
| Contract drafting & review | Critique & Revision | Drafter + Legal Reviewer + Reviser | £18k–£35k |
Single Agent vs Multi-Agent: When to Choose Each
| Choose Single Agent When… | Choose Multi-Agent When… |
|---|---|
| Task fits in one context window | Task volume exceeds context window |
| Steps are sequential and dependent | Steps can run in parallel |
| One set of tools covers the whole task | Different tools needed for different stages |
| Speed of response matters (low latency needed) | Throughput matters more than latency |
| Budget is limited | Quality justifies the extra investment |
| You're building an MVP | Single agent has already proven insufficient |