Multiple AI Agents Working Together

A multi-agent system has roles: a planner breaks down the work, a coder writes it, a reviewer checks it. Each agent has a focused prompt, smaller context, and a clear job. It's how systems like CrewAI, AutoGen, and LangGraph work — and how teams of humans work too.

Some examples

• A research crew: searcher finds sources, reader summarizes them, writer drafts a report.
• A code crew: planner makes a TODO list, coder implements one item at a time, tester runs the suite.
• A content crew: ideator brainstorms, writer drafts, editor cuts.
• A customer support crew: classifier routes tickets, responder drafts replies, supervisor reviews.

Try it!

Sketch a multi-agent system for a task you do weekly. Name 3 roles, each with one sentence describing their job.

One agent that plans and one that executes is more reliable than one agent doing both. The planner writes the brief, the worker runs it — and you get cleaner outputs because each has one job.

Some examples

• A Claude planner outputs a JSON brief that a GPT-4 worker turns into code.
• A research agent writes a one-page summary that a writer agent expands into a blog post.
• A spec-writing agent produces tickets that a coding agent picks up one at a time.
• A triage agent labels emails and a response agent drafts replies using only the labels.

Try it!

Write two prompts — one that plans, one that executes. Wire them so the planner's output is the worker's input.

Multi-agent setups split a task across specialized roles, each with its own prompt and personality.

Some examples

• Drafter writes a first attempt.
• Critic finds three problems.
• Editor decides what to change and produces v2.

Try it!

Set up a 3-agent team to write your next essay outline. Compare to a single-agent version.

Understanding "Multi-agent: a team of 3 agents that disagree productively" in practice: AI agents don't just answer questions — they can do things, like looking things up, writing files, or talking to apps. One agent that drafts, one that critiques, one that decides — better than a single agent — and knowing how to apply this gives you a concrete advantage.

• Design clear agent goals before adding tools
• Define permissions and scope before deploying any agent
• Build in human-approval checkpoints for high-stakes actions
• Understand when to use an agent vs. a simple chat prompt

1. 1Design an agent spec: goal, tools, permissions, stop condition
2. 2Run a simple web-search agent in a sandbox environment
3. 3Instrument an existing workflow to identify where an agent could save time

Frontier AI work increasingly involves multiple specialized agents — a researcher, a writer, a critic, a coder — coordinated by an orchestrator. This pattern produces better work than one model doing everything, but coordination introduces new failure modes. Knowing the pattern helps you spot what's happening behind the scenes in modern AI products.

Some examples

• Research products often run a search agent, an extract agent, and a synthesis agent.
• Coding products may have planner, implementer, and reviewer agents.
• 'Auto' modes in chat tools often spin up multiple specialist calls under the hood.
• When agents pass work to each other, context can get lost or distorted.

Try it!

Next research task, run the search and the synthesis as two separate AI sessions. Compare to one combined session.