Uber’s Multi-Agent SDK | Jamieson Leibovitch, Uber

Uber’s Multi-Agent SDK | Jamieson Leibovitch, Uber

🎙 Jamieson Leibovitch 👥 5K 📅 October 20, 2025 ⏱ 29 min 👁 272 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

AgentFxagent protocolLangGraphMCPeval

Summary

Jamieson Leibovitch, Senior Software Engineer at Uber, presents Uber’s multi-agent platform built on their Michelangelo ML platform. He outlines a five-level autonomy framework for agents, from L1 responders to L5 AGI, and details Uber’s evolution from simple chatbots to advanced multi-agent systems. The talk covers the AgentFx SDK, which provides standard interfaces for building and serving agents, and the Agent Protocol, a fork of LangGraph Cloud APIs, enabling standardized online services. Uber’s platform includes Agent Builder (no-code), Agent Studio (debugging/visualization), and an MCP gateway for integrating internal services. Key features include eval frameworks, guardrails, long-term memory, and support for multiple orchestration frameworks like LangGraph and CrewAI. The speaker discusses cost management, model routing, and governance, emphasizing a thin layer over orchestration frameworks to leverage Uber’s infrastructure. The talk concludes with Q&A on eval, user personas, and open-sourcing plans.

138 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into Uber’s production multi-agent architecture, offering a rare behind-the-scenes look at enterprise-scale agent deployment. The speaker’s argumentation is coherent, grounded in practical experience, and supported by concrete examples like the text-to-SQL project. He effectively explains the rationale behind design choices, such as adopting the Agent Protocol and maintaining a thin SDK layer. The presentation is persuasive, though some claims are high-level and lack deep technical detail, which is typical for a conference talk.

Scientific Rigor, Source Quality, Title Accuracy

The talk is based on the speaker’s direct experience and internal Uber projects, which lends credibility but limits external verification. No specific external sources are cited, only the conference link. The title accurately reflects the content, focusing on Uber’s multi-agent SDK. The presentation is well-structured and aligns with industry trends, but the lack of citations and the proprietary nature of the systems mean the information should be considered as expert opinion rather than peer-reviewed research.

168 words

Title / Content Match

The title accurately reflects the content, focusing on Uber's multi-agent SDK and its architecture.

Quality & Reliability

7/10

The talk is an expert opinion from a senior engineer at Uber, providing detailed insights into their internal agent platform. It is not peer-reviewed but offers practical, first-hand information about production systems. The claims are plausible and align with industry trends, though some details are high-level and not independently verifiable.

Key Moments

Cited Sources

  • MLOps World — Conference website where the talk was presented

Concurring Sources

Contribution & Novelties

The talk provides a unique insider perspective on Uber’s multi-agent platform, detailing the AgentFx SDK and its integration with the Agent Protocol. It offers practical lessons on scaling agents in production, including eval, guardrails, and cost management. The presentation of Uber’s autonomy level framework for agents is a novel contribution to the discourse on agent maturity.

Pour aller plus loin :

  • Agent Protocol — Official specification for agent communication, relevant to the protocol discussed.
  • LangGraph — Orchestration framework used by Uber, relevant to the technical stack.
  • Model Context Protocol (MCP) — Standard for tool integration, relevant to Uber’s MCP gateway.

100 words

Radar Profile

The radar profile shows high scores in information quantity and technical level, indicating a dense, technical talk. The fiabilite score is moderate, reflecting the expert opinion nature. The overall balance suggests a valuable but not fully verifiable presentation.

Reliability 7/10

💬 No comments were provided for analysis.