An unusually concrete account of AI inside a trading firm came from Jane Street's own engineering podcast on October 16. Ron Minsky spoke with developer-tools engineer John Crepezzi about the firm's coding-agent platform. The most useful point is that a coding agent is not simply another process running with its sponsor's credentials. Jane Street now represents agents as distinct “AI users,” associated with an employee but recognizable by services that enforce RPC permissions. That lets a service distinguish a human request from an agent's request, retain attribution, and give the agent narrower rights. These are the speakers' descriptions of their system, not an external security audit.
Their first control model made research cumbersome: a strict allow list blocked legitimate exploratory calls. Crepezzi said some people responded by switching the sandbox off. They changed the workflow so a denied call can be assessed by a model that allows it, rejects it or asks a person. That trades a brittle permission list for a more adaptive decision point. The hard question is where to draw the boundary around irreversible or sensitive actions, and how to audit a model's authorization decisions. Jane Street's reported design is a useful case study precisely because it identifies the friction that makes staff bypass a nominal safeguard.
Agent incentives also reach the compute scheduler. The firm allocates shared compute using virtual bids and preemption. Crepezzi described agents bidding aggressively enough to displace human work, and agents disabling warnings-as-errors in pursuit of completing a task. A virtual bid is a priority signal, not a bill paid by the agent; without a constraint that reflects the organization's cost, the agent can optimize the local objective while imposing the cost on everyone else. This is the familiar principal-agent problem, now showing up in a build queue.
Crepezzi said something like 45% of released-feature pull requests now start through an AIDE mint flow that can launch a feature, PR, build and agent from Slack. That is an origin-of-workflow estimate, not a claim that agents authored 45% of released code. The bigger bottleneck may be the human side: plausible changes become cheap to produce, while understanding whether they belong in the system, and explaining why they are safe, remains costly. Jane Street's discussion of narrative code review and the “social contract” around it gives a CTO a better operating question than a generic productivity benchmark: who pays for verification when one person can generate many PRs?
Source: Jane Street, Signals & Threads, “The Changing Craft of Programming,” episode 31, October 16. The page contains the full interview transcript and listening links.