Events / Minneapolis

Rapid Development of GPT Realtime 2 Browser Agents with actions.json

Technical talk at Lab651 in St. Paul by Yaniv Ben-Ami on building GPT Realtime 2 browser agents using actions.json, his open-source tool. Includes a live Trello demo and Q&A, then networking at Lake Monster Brewing.

Thu, Jul 9, 11:00 PMLab651 · St Paul · MN

Why it matters

actions.json solves a real scaffolding problem: how to declare website actions for a voice-driven realtime agent without hand-wiring every browser interaction. Useful for builders shipping voice agent products.

The tokenmaxxing angle

Realtime voice agents that operate browsers chain audio, reasoning, and browser action tokens in a single session. The actions.json approach reduces ambiguous back-and-forth, which cuts the token overhead of browser agent loops.

From the organizers

Speaker Yaniv Ben-Ami built actions.json as open-source and also works on Babel3, described as an open-source voice-first coding environment for talking to Claude Code. The Trello site is used as the live demo target.