Day 2 — Patterns & hands-on
Goals for the day
- Move from single-agent to multi-agent patterns and know when to use which
- Wire agents to real systems via MCP
- Make agents stateful with memory; make changes safe with evals
Agenda
| Time | Session | Format |
|---|---|---|
| 09:30 – 09:45 | Day 1 recap, parking-lot review | Discussion |
| 09:45 – 11:00 | Multi-agent patterns: supervisor, swarm, pipeline, parallel | Lecture + demo |
| 11:00 – 11:15 | Break | — |
| 11:15 – 12:30 | Lab 2 — build a supervisor agent that delegates to specialists | Hands-on |
| 12:30 – 13:30 | Lunch | — |
| 13:30 – 14:30 | Memory: short-term, long-term, retrieval, compaction | Lecture + demo |
| 14:30 – 15:30 | MCP deep-dive — write a custom MCP server, connect it | Lecture + hands-on |
| 15:30 – 15:45 | Break | — |
| 15:45 – 16:45 | Evals — why they matter more than tests for agents | Lecture + lab |
| 16:45 – 17:00 | Day 2 retro | Round-the-room |
What participants will leave with
- A multi-agent system (supervisor + 2 specialists) running locally
- A working custom MCP server they wrote
- A first eval suite for one of their agents
Materials
- Lab 2 starter repo: link TBD
- MCP server template: link TBD
- Eval scaffolding: link TBD
Attendees
- NBS: TBD
- MetaWeave: TBD
Decisions
- Captured live.
Action items
- Owner — action — due