One agent researches. One agent writes. Together they produce something neither could alone — and the handoff between them is where all the interesting engineering happens. In this lab you design the architecture, sequence the message flow, and watch two real AI agents collaborate live.
Before you build a team, you need to understand the three problems that make single-agent systems break down — and why splitting responsibilities between specialized agents fixes each one.
When a single agent both researches AND writes, it juggles two competing cognitive modes in one context window. Research mode favors breadth — collect everything, stay neutral, flag contradictions. Writing mode favors selectivity — pick the best angle, build narrative momentum, ignore tangents. These modes contaminate each other. The result: a half-researched summary, or a paralyzed researcher who can't commit to a narrative.
Ask a single agent to "research the topic and write a summary" and it will start writing before it has finished researching — because the instruction contains implicit pressure to produce visible output. The research phase gets cut short. Alternatively, it over-researches and delivers a data dump instead of a summary. The two tasks have different completion signals and the agent can't satisfy both.
When research and writing happen inside one agent, you can't inspect what the agent "knew" before it started writing. If the summary is wrong, was it a bad research phase or a bad synthesis phase? In a two-agent system, the handoff message is a first-class artifact — you can read it, validate it, and even route it to a human reviewer before the Writer Agent sees it.
12 responsibilities and capabilities. Each belongs to exactly one role: the Research Agent, the Writer Agent, or the Orchestrator. Drag each to the correct zone — one wrong placement reveals a design flaw.
A two-agent pipeline has 8 distinct steps from user request to final output. Drag them into the correct execution order — pay close attention to when the handoff message gets created, validated, and consumed.
This is real. Pick a topic, hit Run — and watch an actual Research Agent and Writer Agent exchange a structured handoff message and produce a summary. Both agents are powered by Claude.
Not everything in the final summary comes from the research. Some parts are synthesis. Some are the Writer Agent's own inferences. Some may be things the Research Agent never found. Understanding the difference is critical for knowing what you can and can't trust.
Multi-agent systems fail in ways that single-agent systems don't. Classify each failure: which agent caused it, and which stage of the pipeline broke?
Questions on multi-agent architecture, message flow, and the failure modes you just classified. 70% to pass.