Skip to content

Facilitator notes ​

These notes are for whoever is running the room. The goal: keep the day moving, keep energy up, and make sure nobody gets stuck for a whole hour.

The rhythm to protect ​

Each part: ~10 min knowledge-share → ~40 min hands-on → ~10 min checkpoint & reset. Protect the checkpoint at the end, that's the safety net. Announce the baseline switch before anyone falls behind.

Timing guide ​

The times below are a planning aid only, they are approximate durations for each part. Follow the sequence of parts, not the clock; if a part runs long or short, keep the order and let the wall-clock drift.

Approx.PartSegmentWatch for
~45 min1Framing & first agentSetup stragglers; get everyone to one working agent.
~35 min2Context engineeringTreating context as "more is better"; push to what the model should and shouldn't see.
~35 min3MemoryConfusing retrieval with memory; reinforce the retrieval-vs-memory distinction.
LunchEncourage baseline catch-up over lunch for anyone behind.
~35 min4Writing useful specs & personasVague requirements; push to what/why/for-whom/how-long.
~35 min5MCP fundamentalsEverything-is-a-tool tendency; steer read context to resources, workflows to prompts.
~30 min6Build the MCP serverIncomplete tool contracts; attendees skipping the failed-call test.
Coffee breakLast chance to catch up to the Part 6 baseline before the agent parts.
~45 min7Agent Framework & first MCP-connected agentUngrounded answers; agents claiming they acted instead of proposing.
~30 min8Native tools & A2A handoffHave attendees name one native-vs-MCP reason; make sure the failure handoff is run, not just the happy path.
~30 min9Human-in-the-loop & condition checkerMake sure the escalation path is run, not just the happy path; leave 10+ min for the day recap.
ClosePoint to References for continued learning.

Energy checks ​

  • Right after lunch is the low-energy slot, open with a quick demo or a question to the room before the hands-on.
  • Use pair-ups when energy dips: have a working attendee help a stuck neighbor.
  • Call out wins out loud ("three people just got A2A working") to keep momentum.

Room checks ​

Before each hands-on block, do a quick show-of-hands:

  1. "Who's at the checkpoint?", gauge how many can move on.
  2. "Who wants to switch to the baseline?", normalize switching early, not at minute 59.
  3. "Who's on the observe path?", make sure they have the diagrams and checkpoints to follow.

Helping stragglers without stalling the room ​

  • Triage by bucket. Most issues map to Troubleshooting, send them there first before deep debugging.
  • Time-box help to ~3 minutes. If it's not fixed, switch them to the baseline and move on; revisit at the break.
  • Protect the majority. Don't hold the whole room for one machine, offer observe mode and circle back.
  • Environment failures aren't their fault. If a baseline won't run, it's almost always environment-related; move them to observe and rejoin at the next checkpoint.

Scope reminders ​

Keep the room on the default stack (GitHub, Azure, Microsoft Foundry, Python). When someone asks about a different framework, point them to the adapt path in Defaults & paths and keep the room moving.