OpenAI, Anthropic and Two Sigma on financial-agent autonomy and safety
At the AI Engineer New York conference, Joseph Kim, OpenAI's head of ChatGPT for Financial Services, described a "threshold of usefulness" — the point where a model becomes capable and reliable enough to save time — and predicted that by 2027 agents will handle investment idea generation and trading. Two Sigma SVP of Architecture John Feminella outlined five composable controls, arguing safety and innovation are not opposites and criticizing "approval fatigue" as both an innovation drag and a risk. Anthropic's Belinda Neal and Keon Shahab described a three-part trust layer covering agent identity and permissions, sandboxing and approvals, and ongoing validation, while Bloomberg's Andrey Rybka recommended deterministic controls outside the model such as scoped credentials, tool-call controls, outbound restrictions, audit trails, budgets and kill switches.
The talks suggest that deploying agents in regulated sectors like finance now hinges less on model capability than on control design — permissions, approvals and auditability.