Beyond Code Generation: Rethinking Engineering Productivity in the Age of AI Agents

The adoption of AI coding tools has dramatically increased the speed of code generation, but this has created new bottlenecks in review queues, CI systems, verification workflows, and deployment processes. Beyond simply writing code faster, a systemic response is now needed that can safely absorb and verify the increased volume of work.

Dropbox is evolving away from the Copilot model, a simple assistive tool, toward an Agent-centered system that performs work on its own. To carry out a given task, the Agent repeats a cycle of investigating the codebase, editing files, running tests, and fixing issues itself when tests fail.

Nova, an internal platform built to support this, helps AI agents execute tasks in a controlled environment when engineers give instructions in natural language. Nova currently handles about 1/12 of all Pull Requests (PRs) at Dropbox, and beyond feature development, it is actively used for repetitive and labor-intensive tasks such as migrations, test fixes, and bug investigations.

This shift is leading to a change in how productivity is measured. Moving away from the existing approach of simply measuring PR throughput, engineering productivity is being rethought around Product Velocity, which indicates how much of the output generated by AI actually translates into real customer value.