Multi-Agent Collaboration Challenge

Setup

cp .env.example .env   # add your ANTHROPIC_API_KEY
npm install
npm start

This starts a chat channel with one agent (Alex, a software engineer). Type messages and Alex will respond. It works great — with one agent.

Your Challenge

Modify this project so that 3+ agents with different roles collaborate effectively in the same channel.

For example, you might add:

• A product manager who breaks down requirements
• A code reviewer who critiques proposals
• A QA engineer who thinks about edge cases

Requirements

Your system should:

1. Not produce infinite conversation loops — agents should not endlessly respond to each other
2. Not duplicate work — multiple agents should not do the same thing in parallel
3. Have agents contribute based on their role — a reviewer should review, not implement
4. Scale gracefully — adding a 4th or 5th agent should not break things

Rules

• You may modify any file
• You may add new files
• You may change the architecture entirely
• You may use any npm packages
• There is no single correct answer — we care about your design thinking and the trade-offs you make

What You're Given

src/
  types.ts      # Message and AgentConfig types
  channel.ts    # Shared message bus — post, subscribe, getHistory
  llm.ts        # Thin Anthropic SDK wrapper
  agent.ts      # Single agent: listens to channel, responds via LLM
  main.ts       # Boots 1 agent, accepts human input via CLI

The architecture is intentionally minimal. The Channel broadcasts every message to every subscriber (except the sender). The Agent responds to every message it receives. This works fine with one agent.

With multiple agents, you'll quickly discover why this is a hard problem.

Hints (read only if stuck)

What goes wrong with naive multi-agent?

If you just create 3 agents and subscribe them all to the channel:

• Agent A responds to the human
• Agent B sees A's response and responds
• Agent C sees B's response and responds
• Agent A sees C's response and responds
• ... infinite loop

Possible approaches (there are many more)

• Have each agent decide whether a message is relevant to its role before responding
• Introduce a coordinator that decides who should speak
• Add a cooldown or turn-taking mechanism
• Let agents address each other explicitly (@name)
• Set response budgets per "round"
• Build a task queue where agents claim work

Each approach has trade-offs. That's the point.

Time Expectation

This is designed for a 2-3 hour working session. A working (even if imperfect) solution that demonstrates clear thinking about the coordination problem is better than a perfect but unfinished architecture.