Introducing Nova, an Internal Platform for Coding Agents

Dropbox built an internal platform for coding agents, Nova, tailored to its large monorepo, Bazel-based build/test system, and internal infrastructure. The key idea is to treat agents not as simple code generators, but as execution services that run in real verification environments.

Nova runs isolated sessions on a codebase snapshot at a specific commit, executes verification commands as needed, and feeds failure results back to the agent to continue making fixes. It supports a web UI, CLI, and API, and also provides skills, plugins, MCP integration, and observability and evaluation tools.

The use cases are wide-ranging.

• CI failure fixes: When verification fails, the agent analyzes the cause and keeps refining the patch.
• Flaky test remediation: Deflaker sends success/failure logs of flaky tests caught by Athena to Nova, which fixes them after more than 100 rounds of CI verification.
• Migration / dependency upgrade: Nova complements the limitations of the existing Goose-based migrator, and integrates with RenovateBot to prioritize fixing breakages from dependency upgrades.
• Operational experiments: The platform is being expanded into workflows such as reproducing production crashes, making PR review policy decisions, and easing on-call toil.

The lesson Dropbox has learned is clear. An agent's usefulness is determined not just by model performance, but by the design of the surrounding platform, including context, verification, and guardrails — and it's often more stable to manage some steps deterministically outside the agent loop.