Crossing the Threshold: How OpenAI and Anthropic Approach Autonomy and Safety in Financial Agents
Latent Space Richard MacManus
OpenAI, Anthropic, Two Sigma, and Bloomberg all said the same thing: let agents act, but put the brakes outside the model. That’s now the real fight in finance, not raw capability.
Based on reporting by Latent Space, Richard MacManus — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
At AI Engineer New York, the loudest theme wasn’t model size or shiny demos. It was control: how much autonomy to give agents, and where to put the guardrails so they can do useful work without turning into expensive chaos.
OpenAI’s Joseph Kim, head of ChatGPT for Financial Services, called this the “threshold of usefulness” — the point where a model is reliable enough to save time. He argued that rising model capability will push financial analysts past that threshold and into more delegation, and he went as far as predicting that by 2027 the industry will rely less on Excel and PowerPoint and use agents for investment idea generation and trading.
Kim also framed traceability as part of the product, not an afterthought. In finance, people need to see where a number came from and what context surrounds it, so the review step has to feel native to the workflow. And when humans stay in the loop, he said that is a product design choice: what do users still need to control, review, or approve?
Two Sigma’s John Feminella took the same problem and attacked it from the operations side. He warned about approval fatigue, where people are forced to sign off on everything and end up slowing progress or missing important calls. His answer was a set of composable controls, including auto-mode classifiers that decide on routine requests and route only the important ones to a human. In his view, classifiers fit ambiguous cases, while deterministic rules are better for clear ones.
Anthropic’s version of the same argument was about building a trust layer around agents. Belinda Neal said the company partners with financial firms because models cross thresholds and change work. Keon Shahab broke that trust layer into identity and permissions, sandboxing and approvals, and making sure systems still work when variables change. Bloomberg’s Andrey Rybka was even more blunt: keep deterministic controls outside the model, check every call, log it, and keep a kill switch. The shared message was simple. Let agents move fast, but don’t confuse autonomy with handing them the keys.
My take — AI-written commentary, not fact-checked reporting
The industry is finally admitting the obvious: the model is not the safety plan. That sounds boring until an agent reads a review like a command and starts making life interesting in the worst way. The grown-up move is controls outside the model, because “trust us” has never been a great security architecture.
Read more about this at: Latent Space