The problem
Every family asks it three times a day. A fluent suggestion isn't enough: the place has to be open, close enough, and right for everyone eating.
Why I have it
Bourdain answers from places it already knows, with a source and a date for each fact, and searches live only when that comes up short. It's also the first agent on the harness, the test bed for how every agent will be built.
What it does
- Picks places, then checks each pick against its evidence in code.
- Shows its sources with every answer.
- Works under a spending brake on every ask: at most 12 tool calls, 10 searches and $0.50.
Where it helped
- Went live the night before a family trip to New York and Boston, and ran the trip. That trip became its test: the series tells the story.
- 326 places and 3,397 facts, each with its source and the date it was read (Oct 14).
- A judge has graded 214 answers, all in shadow. It gets no vote until it agrees with my own labels.
How it works
One message, step by step
Each step lights up in turn. The names are the real components in the code.
- TelegramIts own bot.
- UnderstandThe ask becomes a typed request. No pattern-matching on free text.
- ContractThe outcomes this answer must meet.
- ExecuteThe knowledge base and maps, under the spending brake.
- VerifyCode checks every pick against its evidence.
- JudgeA second model grades taste. In shadow for now.
- ControlOne repair, or it says plainly what failed.
- Back to meWith a sources line, and the run is recorded.
Not built yet
- Answering from the database first, before any model, isn't built.
- The judge has no authority yet: it needs 50 of my labels and 80% agreement.
- It joins Miya when it passes. It hasn't yet.