Product updates and notes from the MyDataTalk team.
August 6, 2026
Every lever for reducing token spend also removes information the model was using for something. Here's where we actually cut cost in a chat-with-your-database product, and the one place cutting naively breaks the answer instead of just making it cheaper.
Read more →July 31, 2026
A harness governs one model call. A loop is what happens when the agent gets to call again, and again, based on what it just saw. Designing that cycle — its evidence, its termination, its honest failure — is a separate discipline. Here's what we've learned building ours.
Read more →July 25, 2026
Production-grade agents aren't won with a better prompt. They're won in the harness: the constraints, retries, context assembly, and verification wrapped around every model call. Here's how we build ours.
Read more →July 24, 2026
Notes on building an AI data analyst that eliminates the wait between a business question and an answer.
Read more →July 24, 2026
A look at how retrieval-grounded context (not fine-tuning) keeps AI-generated answers honest about your business, and honest about what it doesn't know.
Read more →