Introducing Nova, Our Internal Platform for Coding Agents

Why It Matters

Nova turns repetitive engineering tasks into automated, validated workflows, boosting developer productivity at scale. The platform demonstrates how AI can be safely embedded in enterprise‑grade codebases, setting a blueprint for other large tech firms.

Key Takeaways

• Nova runs AI agents in isolated, commit‑specific environments
• Integrated validation loops keep agent output trustworthy
• Supports both interactive web UI and CLI/API for parallel jobs
• Used to remediate flaky tests with 100+ CI runs
• Enables AI‑assisted migrations and dependency upgrades across thousands of repos

Pulse Analysis

Dropbox’s Nova platform marks a strategic shift from ad‑hoc AI tools to a unified service that embeds coding agents directly into the company’s monorepo and CI pipeline. By snapshotting a specific commit and running agents in hermetic Bazel environments, Nova ensures that any generated code is immediately vetted against the same build and test constraints engineers rely on daily. This tight coupling eliminates the disconnect that often plagues third‑party agents, which operate in isolation from a company’s unique repository structure and validation pathways.

The platform’s design balances autonomy and control. While agents can propose fixes, Nova’s validation commands—such as running targeted Bazel tests—provide deterministic feedback loops. If a change fails, the system feeds results back to the agent for iterative refinement, preventing endless CI stalls and preserving branch hygiene. This approach has already proven effective in high‑impact use cases like flaky‑test remediation, where Nova processes logs from Dropbox’s Athena system, iterates up to five times, and only publishes fixes that survive extensive re‑testing.

Beyond immediate productivity gains, Nova establishes a reusable foundation for future AI‑driven engineering workflows. By exposing a web UI, CLI, and API, teams can launch parallel agents for migrations, dependency upgrades, or even automated incident response. The platform’s observability, prompt evaluation, and feedback mechanisms create a data‑rich environment for continuously improving agent performance. As AI models evolve, Nova’s modular architecture will allow Dropbox—and potentially other enterprises—to scale AI assistance across the entire software development lifecycle while maintaining the reliability and governance required in production environments.