Draft. Outline to replace with the real post:

The two problems nobody solves together

• “Actually useful”: the agent must not silently do the wrong thing.
• “Scalable”: every session costs real inference money; most users never pay.

What a single turn really costs

• Walk one user journey with real numbers: user logs a meal, what the turn costs, what a typical user costs per month, what they pay.

Where usefulness breaks

• A war story: the time the agent committed the wrong food, and what it took to catch it (evals, canaries, tool gates).

Where the money breaks

• Cost per user vs. price paid; why one bad retention month erases the margin.

What I’d tell someone starting today