How Agents Learn From Their Own Runs · Anant Agarwal (TrueFoundry) · Stranger Builds

​Easy mode: read the logs, tweak the prompt, ship, repeat.

​Hard mode: an agent that does that loop by itself.

​

​Anant Agarwal (Principal Applied Scientist, TrueFoundry) builds one live. A basic TrueForge agent runs, diagnoses what went wrong, rewrites its own instructions and beats run 1. Then the part most people skip: how to tell run 2 is really better and not a lucky run.

​

​Along the way he breaks the loop into four parts (practice, checking, updating and diagnosis) and shows how each one breaks, from agents that learn the scorer instead of the job to gains that flatten after a few rounds.

​

​30 minutes: 20 of live demo, 10 of your questions. Nothing to prepare. Building for the TrueForge challenge? Bring the place where your agent gets stuck.

​

​Part of Stranger Builds, NERDCONF's Halloween hackathon. Build week runs until Sun 8 Nov, with USD 10K in prizes and credits. strangerbuilds.com

​

​Anant Agarwal is a Principal Applied Scientist at TrueFoundry, where he leads the customer and product enablement verticals. An IIT Kharagpur and ISB alum, he is an AI educator, conference speaker and Spotify podcast guest.