Citigroup is rolling out Arc, an internal platform that acts as a centralized operating system for AI agents, CTO David Griffiths told Axios in an exclusive published April 30.
Arc gives Citi employees the ability to build autonomous agents that tap into top AI models within a single secure environment. The platform is designed to scale those agents across the firm while maintaining governance and safety controls. It lays the groundwork for what Griffiths described as Citi’s largest push into agentic AI, where multiple autonomous agents orchestrate and complete tasks together.
Developer-First Rollout
Arc will be available to developers first, with plans to extend access to the broader bank over time, according to OODAloop’s summary of the Axios report. The phased approach mirrors how other financial institutions have introduced agent platforms: controlled access, internal governance validation, then wider deployment.
Wall Street’s Agent Infrastructure Race
Citi’s move arrives in a crowded week for enterprise agent platforms. Goldman Sachs made headlines days earlier by restricting Claude access for Hong Kong bankers over geopolitical concerns, while Rogo closed a $160 million Series D for agentic finance workflows and Ramp launched AI agents across its procurement platform.
The pattern across Wall Street is consistent: banks are building internal agent operating systems rather than relying on third-party platforms. Arc joins a growing category alongside Mosaic Singularity’s HeartBeatAgents and Google’s Gemini Enterprise Agent Platform, all announced in the past week.
The Security Calculus
For Citi, the value proposition is routing agents to multiple top models without opening multiple security perimeters. A single envelope means a single audit surface, a single policy layer, and a single set of access controls for agents that may touch portfolio analysis, market trend analysis, and scenario testing.
That architecture reflects the lesson most enterprises learned in 2025: the hard part of agent deployment is governance infrastructure — model selection is secondary.