Descriptions:
Lei Zhang, head of technology infrastructure at Bloomberg, delivers a candid retrospective on what it actually looks like to roll out AI coding tools across one of the world’s largest and most complex engineering organizations—more than 9,000 engineers, one of the largest JavaScript codebases in existence, and a private global network serving the Bloomberg Terminal.
The talk covers Bloomberg’s two-year journey from initial AI tool evaluation to production deployment. Early productivity gains were real but uneven: proof-of-concept generation and test writing improved quickly, while metrics dropped off sharply for anything beyond greenfield work. More revealing is what happened next—average open pull requests increased and time-to-merge slowed, because AI-generated code still required careful human review at scale. Zhang explains how Bloomberg shifted focus from narrow code completion toward broader software engineering tasks, including automated incident response agents that can traverse codebases, telemetry, feature flags, and service dependency graphs without the tunnel vision human responders often bring.
A central theme is platform governance: with teams racing to independently build pull-request review bots and incident response agents, Bloomberg established a structured “paved path” in partnership with its AI organization, including standardized MCP (Model Context Protocol) servers connecting agents to internal metrics, logs, topology systems, and SLOs. This talk is particularly valuable for engineering leaders at large organizations trying to scale AI tooling responsibly without creating redundant, ungoverned systems.
📺 Source: AI Engineer · Published December 16, 2025
🏷️ Format: Workflow Case Study
