Skip to content

05. Assignment 5 — MCP-Powered Briefing Workflow

Week 10. Build one MCP server and one multi-agent workflow that actually justify each other.

Goal

Create a briefing assistant that does four things: 1. researches a topic 2. drafts a brief 3. reviews the draft 4. publishes only approved output

The key requirement is architectural clarity. Do not build “many agents” for drama. Build the minimum multi-agent structure that improves reliability over one giant agent.

What you are building

Part A — MCP server

Expose at least these capabilities: 1. search_reports(query, top_k=5) 2. save_note(title, body, tags) 3. get_editorial_policy() 4. publish_brief(title, summary, body_markdown, review_status)

Requirements: - Use clean tool descriptions - Use explicit schemas or typed inputs - Make publish_brief safe by requiring approved status

Part B — Multi-agent workflow

Build at least three agents: 1. Research agent — gathers claims and sources 2. Writer agent — drafts the brief from validated material 3. Reviewer agent — checks grounding, clarity, and publish readiness

Optional fourth agent: - Publisher agent — formats and invokes the publish tool after approval

Suggested topology

Start with pipeline or orchestrator-worker. Do not start hierarchical unless you can justify it. Do not add debate unless you can measure the gain.

Required deliverables

  1. mcp_server.py or equivalent MCP server entrypoint
  2. agents.py or equivalent workflow implementation
  3. schemas.py or equivalent structured handoff definitions
  4. README.md with architecture diagram
  5. eval_notes.md with failures, fixes, and final verdict on whether multi-agent was worth it

Handoff contract requirement

At minimum, the handoff between agents must include: - task - inputs - constraints - evidence or citations - open risks - done definition

Evaluation requirement

Run at least 10 task scenarios. Track: - end-to-end success rate - one failure example per agent - total latency - total model cost or token estimate

Also compare against a simple single-agent baseline. That comparison is the heart of the hands_on_lab.

What success looks like

  • MCP server works end-to-end
  • Agents use structured handoffs
  • Reviewer catches at least some errors the writer missed
  • You can explain why your chosen topology fits
  • You have a clear answer to: “Did multi-agent actually help?”

Common pitfalls

  • Overstuffed handoffs that recreate one giant context window
  • Vague agent roles
  • Tool definitions that allow invalid actions
  • No baseline comparison
  • Fancy topology with no measurable benefit

Stretch goals

  • Cheap router model before generation
  • Human approval gate before publish
  • Parallel evidence gathering in research
  • Trace visualization or structured logs

Write-up prompts

Answer these in your README or eval_notes.md: 1. Why did you choose this topology? 2. Where did MCP help reuse or clarity? 3. What was the first failing handoff? 4. What did your single-agent baseline do better? 5. Would you keep this design in production?

Cross-reference

Use 02_explainer.md for the mental model. Use 03_study_material.md as the compact checklist. Use 06_revision.md after shipping.