Multi-agent Systems

Multi-agent Systems

A single agent is powerful. A team of agents — each with specialised roles, working in parallel or in sequence — is more powerful still. Multi-agent systems are becoming a standard pattern for tackling complex, long-horizon tasks.

Why Multiple Agents?

A single agent faces practical limits:

Context window: a complex task might generate more reasoning and tool output than fits in one context window.

Parallelisation: some tasks can be broken into independent subtasks that run simultaneously. A single agent must execute sequentially.

Specialisation: different agents can be optimised for different roles — a planner, a coder, a reviewer, a researcher — each with the right tools, system prompt, and model for its task.

Error checking: one agent's output can be independently reviewed by another, catching mistakes before they propagate.

Reliability: if one agent fails, the orchestrator can retry or route around it.

Common Multi-agent Patterns

Orchestrator → Workers: a planning agent breaks a task into subtasks and dispatches them to specialised worker agents. The orchestrator collects results and assembles the final output.

Pipeline: agent A produces output that becomes input for agent B, which passes to agent C. Each agent specialises in one transformation.

Parallel fan-out: a task is sent to multiple agents simultaneously, each attacking it from a different angle. Results are compared or merged.

Critic/reviewer: one agent produces output, another reviews and critiques it, the first revises based on feedback. Iterates until the reviewer is satisfied.

Debate: two agents argue opposing positions, with a judge agent deciding. Useful for complex decisions or stress-testing reasoning.

Challenges and Guardrails

Multi-agent systems amplify both capabilities and failure modes:

Cost: each agent is making API calls. A complex multi-agent system might make dozens or hundreds of calls per task. Budget carefully.

Debugging: when something goes wrong, it can be hard to trace the failure through multiple agent handoffs. Good logging and tracing are essential.

Trust boundaries: should one agent blindly follow instructions from another? A compromised or misbehaving agent in a chain can cascade. Apply the same prompt injection vigilance between agents as between users and agents.

Stopping conditions: agents need clear criteria for when they're done. Without good stopping conditions, agent loops can run indefinitely.

Human oversight: complex, long-running multi-agent tasks benefit from checkpoints where a human can review progress and course-correct.