Multi-Agent System
A system where several specialised agents collaborate — each with a defined role, sharing state via messages or memory.
What is Multi-Agent System?
A multi-agent system splits a hard task across specialised agents. A "researcher" agent gathers information, a "writer" agent drafts, a "critic" agent reviews. Each agent has its own prompt, tools, and possibly its own model — letting you optimise each independently.
Patterns: **supervisor / worker** (one agent coordinates, others execute), **debate** (two agents argue to surface flaws), **pipeline** (sequential handoff), **swarm** (parallel exploration with merging). LangGraph, CrewAI, and Anthropic's patterns paper formalise these.
Multi-agent is not always better than single-agent. It adds latency (multiple LLM calls), cost, and failure modes (agent A waits for agent B forever). The 2026 rule of thumb: try one well-prompted agent first; reach for multi-agent when the task has genuinely independent sub-tasks.
Multi-agent design is the senior-level agent engineering skill. Knowing when NOT to use it is as valuable as knowing how.
A Bengaluru research-tools startup built a 3-agent academic search: a query expander, a parallel searcher across 6 sources, and a synthesiser that produces a cited summary. End-to-end latency 18 seconds; quality beats single-agent baseline by 40%.
Want to master this?
Learn Multi-Agent System in a structured cohort
3-month live program with mentors, real projects, and 50+ partner placement support.