# Multi-Agent Systems: When One Agent Is Not Enough

> Orchestrator–worker, pipeline and debate patterns, what they cost, and how to decide whether you need more than one agent.

- Source: https://akshaythakare.com/notes/agentic-ai/multi-agent-systems
- Topic: Agentic AI
- Level: advanced
- Reading time: 2 min
- Tags: agentic-ai, agents, multi-agent, architecture, interview
- Updated: 2026-10-11

## Start with one agent

A single agent with good tools and clear instructions solves most tasks. Multiple agents add latency, cost, coordination failures and harder debugging. Add them only when they solve a specific problem.

## Good reasons for multiple agents

| Reason | Example |
| --- | --- |
| Independent sub-tasks in parallel | Research ten companies at once |
| Context isolation | Each worker reads its own large documents; the orchestrator sees summaries |
| Different tools or permissions | A read-only research agent and a separate agent that can write to a system |
| Specialised instructions | A reviewer agent that checks another agent's output |

## Common patterns

```text
Orchestrator–worker:   orchestrator → plan → [worker A, worker B, worker C] → combine
Pipeline:              extract agent → validate agent → summarise agent
Generator–critic:      writer → reviewer → writer (until the checks pass or a limit is hit)
```

## Make it reliable

- Pass **structured** handoffs (JSON with fields), not free-form chat.
- Give each agent a narrow role, its own tools and a step limit.
- Keep shared state in code or a database, not only in prompts.
- Log every step with inputs, outputs and token cost so failures can be traced.

> **Tip: Interview framing**
>
> "I start with one agent and add workers only for parallel or isolated sub-tasks, with structured handoffs, step limits and tracing."