Summary
At the AI Engineer conference, Shu Fang from Two Sigma — a 25-year-old quantitative hedge fund — presented how the firm built a system in which every employee runs a personal cloud agent operating under their own user identity rather than a separate machine identity. The talk, titled “Tethered: Our Agents Are Us,” walked through the engineering tradeoffs and the practical reasoning that led Two Sigma to an unconventional but production-tested architecture.
Fang explained that the conventional approach of assigning agents a separate machine identity collapses quickly in enterprise environments: permissions drift out of sync, software licensing doubles, systems like Google Workspace struggle with multiple identities accessing shared data, and audit trails become ambiguous. Two Sigma instead runs agents in per-user Kubernetes namespaces, with a sidecar identity service ensuring all agent actions are cryptographically tied to the human user. Critically, a trace ID is propagated through all agent actions using HTTP clients, MCP, and skill harnesses, enabling full provenance replay across multi-step workflows — something a separate agent identity would make impossible.
For regulated web access, Two Sigma uses Google’s enterprise web grounding product rather than allowing external egress, providing search and fetch capabilities within the company’s existing VPC with data freshness of 6–24 hours. Fang presented the full approach as a practical template for deploying agents safely in highly regulated industries, with concrete takeaways on identity architecture, observability instrumentation, and the security benefits of tethering agent capabilities to human accountability.
📺 Source: AI Engineer · Published September 03, 2026
🏷️ Format: Deep Dive
