We are thrilled to announce the winners of the Sanity Challenge, our partnership with Sanity built around a single idea: an agent is only as good as the content behind it.
Two paths were open. Path one asked you to ship an agent backed by a Sanity Context Knowledge Base, with a bar we set deliberately high: if a keyword search would have gotten the same answer, aim higher. Path two asked you to vibe-code something strange and tell us the truth about how it went.
Narrowing this field down was tough. We saw an agent that knows Romania rewrote its vignette rules this week, and will tell you in four languages, an etiquette guide that refuses to store a verdict and works out whether something is rude from what locals and travelers actually say, and more than a few projects that measured themselves against keyword search and published the rounds they lost. Many of you wrote down every wrong turn instead of just the clean ones, which made reviewing your builds a blast.
Thank you to everyone who participated. Now, let's celebrate our winners!
🎉 Congratulations To…
Path One Winners: Ship an agent that queries real content
@rayyer built an agent that answers what the clock and the calendar legally say at a given place and moment, and on whose authority, reading the law, the IANA time zone database and the runtime's own clock side by side. Pending legislation is modeled as Sanity Content Releases, so you can ask what the answer would be if a bill passed, and when two sources disagree the agent drafts a ruling that a person signs before it becomes a standing instruction for every future build. On 30 held-out questions Clockwrit scored 30/30, against 22/30 for the same model with no tools and 16/30 for keyword search over the same sources.
@dannwaneri studied geophysics at university and built an agent for the problem he was trained to work on: whether a groundwater resistivity reading at a Niger Delta site is normal, anomalous, or uncertain, grounded in three published survey papers. Numbers come only from GROQ queries and explanations only from the Knowledge Base, and because survey sites in the Niger Delta often have no network, an offline mode drops the model entirely and still returns the verdict from the structured data. While building it, he found internal inconsistencies inside those three papers, checked each one against the source PDFs, and contacted the corresponding authors.
@big14way built Clearance Desk for a problem with a one-year cost: in Nigeria you can be offered admission and still be turned away at clearance over a single wrong subject. It checks a candidate's UTME subjects, score and O'level sittings against the published 2026/2027 requirements at five universities and returns Eligible, At risk, or Not eligible with every source linked, on the design rule that the model finds and explains while deterministic code judges. His eval answered our own question directly: the same model with keyword search over the same Knowledge Base read the right entry, quoted the right rule, and still told an ineligible candidate she was fine.
Path Two Winners: Vibe-code something strange
@faris-khan turned 2,512 real NYC 311 pothole complaints from one Queens community board into a pet adoption shelter where every fate follows the city's own record. The gaps in the data became the characters: 62% of complaints have no coordinates, so they are the lost strays, listed by street rather than placed on a map that would have to be invented. Three Workflows definitions run the lifecycle and a custom App SDK app handles moderation. The build writeup is an honest one, covering the document cap migration, the duplicate workflow instances, and the moment the agent reported that nothing had been written when something had.
@kenwalger built a virtual wine cellar that answers a harder question than "what do I have right now," storing events and dated claims rather than current state so you can ask what the collection looked like in 1999, or what you believed a bottle's drinking window was in 2018. The build writeup is a study in making verification hard to fake, with gates that name what does not count as proof and oracles the model was explicitly told not to repair. Then a reader found a flaw in the design after publication, and rather than quietly patch it, Ken added an addendum explaining the catch and recorded it as an architecture decision record.
Prizes
All five winners will receive:
- $500 USD cash prize
- DEV++ Membership
- Exclusive DEV Badge
All participants with a valid submission will receive a completion badge on their DEV profile.
What's Next?
Week 3 of the Hacktoberfest Open-Source AI Challenge is live and the theme is Slow and Steady. Check out the full details and join today:
Week 4 of the Hacktoberfest Open-Source AI Challenge launches on October 26. Register now so you don't miss the theme reveal:
Hacktoberfest Open-Source AI Challenge: Week 4
October 26, 2026Register now to be notified as details drop.
Thank you to Sanity for partnering with us on this one, and congratulations again to our winners 💜