LAB GUIDE / SECTION 08
Rounds
Time to put your Droid in production, with the whole class watching.
How a Round works#
- The teacher presses Go. The Hub sends the same question to every Droid, 6 at a time, so cards turn green in waves, not all at once. Don't panic if yours is last.
- The projector shows every Character card with its status and response time. It shows no answer text, only whether your Droid worked.
- The teacher reads all the answers and features the fun ones on the projector. Featuring isn't ranking: there are no winners.
- Fix Window (≈ 3 min): red card? Find out why, fix it with your agent, review, push.
- The Round runs again. Watch your card turn green.
After a Round closes, you can read every answer at https://arena.cf2026iscte.nobledemos.com/round.
| Round | round | What the Hub sends |
|---|---|---|
| 1 | intro | "Introduce yourself to the class…" |
| 2 | scenario | A random planet, starship or species from SWAPI turned into a challenge, e.g. "Sell the class a holiday on Hoth" |
| 3 | scenario | Another random scenario. Try a new tone before it! |
Reading your status#
| Card shows | Means | Look at |
|---|---|---|
OK 1.4s | Worked, took 1.4 s | Nothing. Enjoy it |
timeout | No answer within 8 s | npx wrangler tail: which call is slow, SWAPI or the AI model? Is max_tokens small? |
500 | Your Worker threw an error | npx wrangler tail (below) |
404 | No /api/chat route, or the last push didn't deploy | Dashboard → Deployments |
invalid JSON | Your Droid didn't return { "answer": "..." } | The return at the end of handleChat |
unreachable | The Hub can't reach your URL | Open https://<your-droid-name>.<your-subdomain>.workers.dev/api/card yourself |
Debug in production: wrangler tail#
npx wrangler tailThis streams live logs from your deployed Worker: every request, every console.log, and every error with its stack trace. Leave it running during the Rounds and you'll see the Hub's request come in.
Tip
This is how real teams debug: no SSH, no server. Just logs from wherever in the world your code ran. You can also see them in the dashboard under your Worker → Logs.
During the Fix Window#
read the error in wrangler tail → paste it into your agent → read the diff → push → wait ~30s → green cardGive your agent the exact error from wrangler tail and ask for one fix. "Here's the error, fix it in src/droid.js" uses far fewer tokens than "my droid is broken". No time to wait for the agent? Fix it by hand. It's your code too.
Ideas between Rounds:
- Answers too long, bland or off-character? Rewrite your
TONEand make it sharper. That one is still yours to write. The 3B model sometimes drifts, and a concrete tone keeps it on track. - Curious? Ask your agent to switch
MODELto another chat model from the Workers AI catalog and compare answers and speed on the next run. Avoid "reasoning" models: with only 120 tokens they spend them all thinking and return nothing, so your Droid falls back.
Check#
- Your card went green at least once
Next: What's next
Checks are saved automatically on this browser.