Dropbox's Nova: A Platform for AI Coding Agents

The Challenge: Integrating AI Agents into a Complex Dev Environment

At Dropbox's scale, integrating AI coding agents effectively required more than off-the-shelf tools. The existing development environment, characterized by a large monorepo, Bazel for builds and tests, and on-premise infrastructure, posed unique challenges. Agents needed to operate within these specific constraints, including custom validation paths and infrastructure dependencies, rather than introducing separate, AI-specific workflows. This drove the decision to build a platform like Nova instead of a collection of single-purpose solutions.

Nova's Architectural Approach

Nova is designed as a shared platform that supports interactive development, background jobs, and internal services, ensuring consistent execution, validation, and context handling. Key architectural decisions include: an isolated execution environment for each session with a snapshot of the codebase, caller-provided tasks and optional validation commands to ground agents in real build/test environments, and support for multiple coding agents behind a unified interface.

{ "repo_commit": "<commit-sha>", "task": "Investigate this CI failure and propose a fix", "validation_commands": [ "bazel test //path/to:test_target", "bazel test //path/to/related:all" ], "continue_on_validation_failure": true, "max_iterations": 5, "push_branch": "ai/nova/ci-fix" }

• Skills, Plugins, and MCP Integrations: Nova provides agents with access to observability systems and other internal tools, allowing them to gather evidence, read logs, and inspect failures beyond just editing files.
• Deterministic Code Publication: Each session is limited to a single branch, keeping code publication outside the agent's direct control. This simplifies workflow automation (e.g., testing, rebasing) and ensures predictability.