The speaker introduces the central problem of how to architect communication between thousands of AI agents operating at a global scale.
AI is evolving from reactive systems to an agentic era where software components plan and execute tasks autonomously to achieve high-level goals.
An AI agent is defined by its persona, perception, short and long-term memory, and the tools it can use to perform actions.
Scaling AI agents presents challenges in state synchronization and observability, making event-driven architecture a necessary paradigm for asynchronous communication.
A pre-recorded demo showcases an Uber-like application with robo-taxis and users interacting asynchronously, powered by an event-driven backend.
The architecture uses Kafka topics for users, taxis, and rides, with an orchestrator agent processing events and a simulator using Mapbox for routing.
A live demonstration spins up containerized services to simulate users requesting rides and robo-taxis being dispatched in real-time in Berlin.
A key lesson learned was to avoid unnecessary LLM calls that increase cost and latency, emphasizing the need for observability and pragmatic design.
The final demo showcases the architecture's scalability by running simultaneous simulations in both Amsterdam and Berlin.
39:11Thomas Dohmke
24:22Philipp Schmid
29:03Jörg Neumann
26:35Ricardo
30:01Mario Fusco
25:14Ivan Burazin
23:33Oliver Will
24:12Brad Axen
Jobs that call for the skills explored in this talk.
DTS Systeme GmbH