DemosGames, a ticket desk and three agents, each deciding with Amos 1.

Every guess, move, route and click below is one request to Amos: a case and a few typed questions in, answers out with how sure it is. The browser agent, the feed filter and the cutoff demo use that to hand unsure steps to a person.

Live

These call the API at every step, so each answer is the model deciding as you watch.

choice · the words still possible
Wordleone call per guess
choice · four directions
Snakeone call per move
248421623286412884
↓←→↑
choice · four swipes
2048one call per swipe
choice · after every stroke
Quick, drawa guess after every stroke
- playground/src/tests/
- amos/
- serve.pydecoder.pydata.py
choice · the next folder down
Code searchone pick per folder
⏺(bun test 2>&1 | tail -80); echo "---"; ls -R src testsallow 0.95⏺Edit src/invoice.tsallow 0.94⏺git push -u origin fix/invoice-single-rounding 2>&1 | tail -5ask 0.88
choice · allow, ask or block
Coding agentauto mode for every call
flyport.example/search?OPO-LIS07:15€8909:40€11213:05€97click 09:40 · 0.95
choice · the next click
Browser agentone call per click
Shipped dark mode today. Took three tries to get the contrast on disabled buttons right.not an ad · 0.96software · 0.92
yes / no · is it an ad
Feed filterads hidden, unclear posts to a person

Recorded runs

Saved runs you can step through, so the same tickets can sit next to other models' answers.