Teamday
← People

John Feminella

AI Architecture and Engineering at Two Sigma

agentsenterprisesafetytwo-sigma

About John Feminella

John Feminella works on AI architecture and engineering at Two Sigma, the New York hedge fund. He spoke at AI Engineer (2026) on how a heavily regulated firm can give AI agents real autonomy without locking everything down, stressing that the views were his own and not Two Sigma's.

His talk, "What a $900M Bank Error Teaches Us About AI Agents," starts from the August 2020 incident in which Citibank, as administrative agent on a Revlon loan, sent lenders about $900 million of principal instead of roughly $8 million of interest, even though warnings fired and three approvers signed off. For Feminella the lesson is that controls fail when everyone follows the procedure without seeing the system as a whole, and that agents need controls designed for that reality.

Career Highlights

  • Works on AI architecture and engineering at Two Sigma
  • Speaker at AI Engineer (2026) on governing AI agents in regulated enterprises

Notable Positions

On Safety and Innovation

Feminella rejects the trade-off between safety and speed. A locked-down sandbox is safe but useless, letting agents run free is too weak for a regulated firm, and policy documents in prompts can't be enforced at runtime. Agents, like human employees, should get boundaries, visibility and controls.

On Five Composable Controls

His framework: an agent registry that records every agent's owner, tools and endpoints automatically; risk scoring so a ticket summarizer doesn't wait two months behind a production deployer; guardrails enforced from each agent's declared purpose; shared, firm-run evals treated like unit tests; and runtime classifiers that handle routine actions so humans only review ambiguous or risky requests.

On Approval Fatigue

Asking people to approve every agent action leads to reflexive "yes" clicks. If a request is approved virtually every time, it should be a rule, not a prompt to a human. See Human-in-the-Loop.

Key Quotes

  • "You can't really know in advance what an agent will do, right?" (on non-deterministic systems)
  • "A sandbox with no real access to systems is safe, but it's not very interesting or useful." (on locking agents down)
  • "Why do cars have brakes? It's not just so that you can stop. It's so that you can go fast and stop safely." (on controls)