The interesting question at AI Engineer New York this week was not how capable agents are, but where a human belongs in the loop. Latent Space’s write-up of the October 14, 2026 sessions gives the two halves of that argument: OpenAI’s Joseph Kim described a “threshold of usefulness” - the point where a model becomes capable and reliable enough to save you time - and Two Sigma’s John Feminella argued that approval fatigue, not model quality, is what actually caps autonomy.
Key facts
| Fact | Detail |
|---|---|
| Question answered | The ‘Threshold of Usefulness’: How OpenAI and Anthropic Design Agent Autonomy |
| Short answer | At AI Engineer New York on October 14, 2026, OpenAI’s head of ChatGPT for Financial Services described a ‘threshold of usefulness’ - the point where a model is capable and reliable enough to save you time - wh… |
| Reading time | ~2 min |
| Last updated | 2026-10-15 |
OpenAI: know when the agent pays for itself
Kim, Head of ChatGPT for Financial Services at OpenAI, framed capability as something you cross rather than accumulate. Below the threshold, an agent costs more attention than it returns; above it, analysts can delegate real work. He predicted that in 2027 the financial industry leans less on Excel and PowerPoint and that agents become good enough for “investment idea generation and trading”.
He also described traceability as a product feature rather than a compliance afterthought: in finance you need to check where a number came from and understand its context. His framing of human-in-the-loop work is worth borrowing verbatim as a design exercise - how would the interface change, and what would the user still need to control, review or approve?
Two Sigma: composable controls beat big restrictions
Feminella, SVP Architecture at the hedge fund Two Sigma, rejected the idea that safety and innovation sit at opposite ends of a spectrum. What he objected to is approval fatigue: a person who must approve everything slows the work down and, because attention is finite, raises risk rather than lowering it.
His answer is a set of composable controls - guardrails and shared capabilities, each doing one job - including auto-mode classifiers that decide routine calls and escalate only the requests most likely to matter. He draws a line between the two mechanisms: classifiers for ambiguous cases, deterministic rules for clear ones.
Anthropic: context first, then the boundary of trust
Anthropic’s session, following last month’s Claude for Financial Advisors, described building trust into financial agentic workflows: give the agent context, then define what it is allowed to do without asking.
What to copy on your own machine
- Write down what your agent may do unattended and what needs a human; most “safety” debates are really that list, unwritten.
- Kill approval fatigue deliberately: route routine steps to automatic handling and reserve prompts for irreversible actions.
- Prefer deterministic rules where the case is clear, and save judgement calls for the model.
- Keep traceability local, so you can see which step produced a number.
If you want an agent that runs on your own Windows desktop with an approval gate you actually understand, download MOOGH.
Related reading: