James Reggio

James Reggio was Brex’s chief technology officer in January 2026, following earlier work on its mobile app and AI capabilities. He helped organize the company’s AI strategy and develop infrastructure for operational automation and customer-facing agents. His consequential technical decisions included moving toward specialist agents that converse with one another and favoring auditable standard operating procedures after an unsuccessful reinforcement-learning experiment in underwriting.[4][5][2]

As CTO, Reggio organized Brex’s AI investments into three pillars: corporate AI for employees’ workflows, operational AI for the work of running a financial institution, and product AI for customer-facing capabilities. In his Latent Space interview, published January 17, 2026, he explained that this framework gave scattered experiments a basis for a vision and roadmap. His own involvement concentrated on operational and product AI; IT and the people team drove much of the corporate-AI effort.[2]

That strategy built on earlier infrastructure. Reggio recalled leading an internal AI labs team before becoming CTO and helping establish an LLM gateway around January 2023, although he was tentative about the month. The platform supported prompt deployment, versioning and evaluation, data egress, model routing, observability and cost monitoring. At interview time, it still powered smaller LLM applications and much of Brex’s operational AI.[2]

For the newer agent layer, Reggio and the team chose TypeScript, building against public interfaces while the existing backend remained in Kotlin or Elixir. They adopted Mastra partly because its ergonomics resembled their internal LLM framework and maintaining that framework was an investment they wanted to reduce. This was a partial migration: Reggio said roughly half of the agent-layer applications used Mastra, while the remainder used another internal framework focused on networks of agents.[2]

Brex’s employee assistant began with a broad agent, many tools and retrieved context. Reggio described its difficulty handling expense management, travel, policy and procurement together. Loading more tools or switching the prompt’s context by topic also proved less effective for their use cases. The team instead built specialist agents behind an orchestrator, allowing agents to exchange messages across multiple turns. This division also let a domain team, such as travel, refine its specialist without making one team responsible for every employee action.[2]

Reggio illustrated the distinction with an employee asking how much they could expense per person for a team dinner. The policy agent might need to know whether it was a customer event, a team event or part of a trip. It could ask the assistant to obtain that clarification from the employee before answering. As he put it, “a tool call is basically just one RPC”; the specialist interaction needed a conversation. He treated MCP and tool calls as interfaces to conventional systems, while agent conversations coordinated the work.[2]

His audit-agent example extended that architecture across finance teams and employees. An audit agent raised potential violations with an emphasis on avoiding false negatives—for example, a pattern of transactions just below a receipt threshold. A separate review agent judged whether the amount, context and employee’s broader compliance behavior warranted a case. Filing a case could then trigger the employee’s assistant to collect a business justification or supply context it already held. These were distinct detection, judgment and follow-up roles, rather than a single agent deciding and pursuing every flag.[2]

Reggio reported that Brex invested in reinforcement learning for credit decisions and underwriting with an outside specialist company, hoping to reproduce a human underwriter’s decisions. The resulting performance was inferior to a web-research agent, he said; he supplied no numerical comparison. His takeaway was to break operational problems into granular standard operating procedures that were repeatable and auditable. For many tasks, simpler agents or even single-turn completions were sufficient. The harder work was making prompts capture the procedure and institutional knowledge that people had not written down.[2]

Reggio credited COO Camilla Matias with helping operations staff move from executing procedures toward building prompts and evaluations. Engineering supported that shift with Retool interfaces for prompt, tool and evaluation management. Operations specialists could refine prompts and test new models without needing engineers for each change, comparing performance, latency and cost. Reggio distinguished accuracy evaluations that blocked unacceptable regressions from subjective tone and coherence measures tracked over time, and described increasing evaluation rigor after shipping.[2]

Reggio described a procurement policy of buying seats across competing AI providers and coding tools rather than choosing one winner. Employees could select and switch their stack through ConductorOne. Brex also redesigned its engineering interview around a project requiring agentic coding, assessing how candidates worked and whether they understood the generated code. Existing engineers, including managers, took the exercise too. Reggio said internal scores were not retained and he had no pass/fail data; the purpose was to prompt practical skill development.[2]

Reggio did not emphasize the percentage of code written by AI as a productivity measure. He was more concerned with review rigor, maintainability and gaps in teammates’ understanding of services changing rapidly under agentic development. He argued that agents amplified poor architecture and misunderstood requirements alongside good outcomes, making capacity gains more nuanced than a simple headcount calculation. In a January 2026 Brex accounting announcement, he expressed a related requirement for financial software: “With accounting, accuracy and speed can’t be trade-offs.”[2][4]

Before this AI leadership work, Reggio had a background in mobile engineering and company formation. The Engineering Leadership Podcast’s November 2025 biography says he joined Brex in 2020 as a principal engineer and contributed to its mobile app and AI capabilities. It lists Microsoft, Salesforce, Airbnb and Stripe among his earlier software-engineering employers. The same biography describes Banter as a podcast social-discovery platform he founded and says Convoy acquired it in 2018.[5]

Reggio attributed his move into the CTO role partly to the leadership and business skills he developed as a founder. He recalled considering leaving Brex to start another company when the CTO offer arrived. That experience also informed his support for hiring former and aspiring founders: he valued giving them substantial problems and immediate customer distribution, while celebrating employees who later left to build companies. He described Brex’s “Quitter’s Welcome” recruiting initiative as an explicit embrace of that approach.[2]