Uber's Multi-Agent Platform
summary
Uber claims single agents aren’t enough and ships a production stack: AgentFx SDK (internal agent runtime with a LangChain-linked agent protocol fork of Langraph), a no-code Uber Agent Builder for event-driven/Kafka agents, and Agent Studio with visualization, OpenTelemetry tracing, offline evals and token cost dashboards. The platform supports MCP, multimodal and OSS models, runs multi-agent graphs with observability and an eval pipeline, counts ~2–3k internal Builder users, is closed source today, and the core team is ~10 people.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker background
- 1:16Levels of agent autonomy
- 4:03Evolution of Uber’s agent platform
- 7:13Agent Protocol and SDK design
- 10:48Agent Builder and Agent Studio demo
- 15:00Current platform capabilities
- 21:52Getting started with the tools
- 26:49Adoption by non-engineers
- 28:15Team size for platform support
- 29:41Use cases across teams
- 31:07Q&A and closing
related talks
30mLessons From Building and Deploying a Multi-Agent System at Scale: An...
20mWorkshop
43mThe AgentVerse - Build fast, Scale smarter
34mUnlock the full potential of Generative AI with Agent Platforms
15mThe Responsible Path to Hyperproductivity with AI Agents
40mBuilding Agents with Agent Bricks and MCP
21mPlatforms for Humans and Machines: Engineering for the Age of Agents
01h 00mBuilding a Polyglot Multi-Agent AI Architecture