Skip to content
Flue Agent Workshop

Finish line

Finish line

Section titled “Finish line”

You built a durable agent as six small, testable behaviors. Run one final conversation that exercises the complete system. Trip dates are filled automatically, reusing your workshop dates:

Terminal window
npm run deploy
sleep 5
npm run smoke -- https://field-trip-agent.<subdomain>.workers.dev final-demo \
  "Plan a two-day offsite in Lisbon from <START> to <END> for 12 people. We like architecture, local food, and the outdoors. Keep it under 500 EUR per person."
TIMEOUT_S=200 npm run smoke -- https://field-trip-agent.<subdomain>.workers.dev final-demo \
  "Check the weather, research three suitable places, then write and read itinerary.md."

Save your workers.dev subdomain to open this test ↗https://field-trip-agent.<subdomain>.workers.dev/?id=final-demo

This opens the same conversation as the smoke test. If you repeat this checkpoint, before copying its commands.

To repeat the final test interactively, open the chat link for your selected tab:

  1. Select New conversation.
  2. Send Plan a two-day offsite in Lisbon from <START> to <END> for 12 people. We like architecture, local food, and the outdoors. Keep it under 500 EUR per person.
  3. In the same conversation, send Check the weather, research three suitable places, then write and read itinerary.md.

Verification gate

A complete run should show

  1. The trip brief is saved and available on the next model turn.
  2. Weather tools call Open-Meteo.
  3. The parent delegates venue research to venue-scout.
  4. Cloudflare Sandbox writes and reads itinerary.md.
  5. Agent tracing records the complete turn under invoke_agent FieldTrip.
  6. Session replay and the trace waterfall show model, tool, subagent, and Sandbox work.

Inspect the final conversation in the Cloudflare Agents dashboard. Its trace should contain model and tool spans, venue-scout delegation, and Sandbox work.

Clean up when you are finished

Section titled “Clean up when you are finished”

Your workers.dev endpoint is public and can invoke billable AI and container work. Keep it only if you plan to continue experimenting. Otherwise, in your own workshop account, list and delete the field-trip-agent-sandbox application, then delete the Worker:

Terminal window
npx wrangler containers list
APPLICATION_ID="paste-the-field-trip-agent-sandbox-application-id"
npx wrangler containers delete "$APPLICATION_ID"
npx wrangler delete field-trip-agent

For local-only cleanup, stop Vite and remove .wrangler/state. These actions delete workshop state and cannot be undone.

Six ideas worth keeping

Section titled “Six ideas worth keeping”
  1. An agent is a function that re-renders before each model call.
  2. A conversation ID maps to a durable, isolated state boundary.
  3. Tool schemas are contracts between probabilistic reasoning and deterministic code.
  4. Subagents need complete briefs because they start with fresh context.
  5. Sandbox files and typed persistent state have different lifecycle guarantees.
  6. Agent traces make multi-turn, multi-tool behavior explainable.
Take it with you

Download the workshop deck

The PDF contains the concise Flue introduction, field-day mission brief, and workshop takeaways used during the live session.

Download PDF Open interactive deck

Keep experimenting

Section titled “Keep experimenting”

Finished early? Continue with the six guided bonus assignments, including an on-demand workspace, a named AI Gateway with live Guardrails, model telemetry, failure recovery, a model comparison, and content-free traces.