An agent is only as good as its model of reality, and reality moves. That gap is where a lot of the most interesting work is happening, across both research and production.
A few threads worth building on:
Keeping a live world model. The earliest versions of this idea - systems that watch the world and act when a condition is met - go back a decade, but they were narrow and hand-specified. Capable models are generalizing them into always-on agents that hold a current picture of a market, a competitive landscape, or a regulatory environment, and act the moment it shifts.
Grounding. The path to agents you can actually trust runs through verifiable, current information: attribution, re-verification, retrieval that reflects today rather than a memorized snapshot. Closing the gap between fluent and correct is one of the field's most valuable open problems.
Acting, not just reading. The frontier moved from agents that summarize a page to agents that operate on the live web — comparing, booking, and transacting against real prices and real inventory, the agentic-commerce wave now taking shape. This is where the demo-to-production gap is widest and most worth closing.
The web as a proving ground. Static benchmarks get memorized and go stale; live data opens the door to evaluation and training on real, fresh tasks — contamination-resistant evals, and the live web as an environment where agents can act, learn, and self-improve.This evening is a few hours to build at that edge with people working on the same problems.
FORMAT6:00 PM — Doors open6:30 PM — Lightning talks6:45 PM — Build session9:15 PM — Demo presentations9:30 PM — Close
SPEAKERS Hao Zhu: Founder and AI researcher focused on cooperative AI agents and human-agent interaction
TECH STACKBright Data provides API access, documentation, and starter templates for all participants.Register via: https://app.agihouse.org/events/real-time-agents-build-evening
