Dropbox

Dropbox wires MCP and its Dash content index into code review so an LLM can retrieve the original threat model for a feature and judge whether the diff implements the agreed mitigations — closing the months-long gap between security design and shipped code.

Dropbox's rebuilt API documentation adds interactive in-browser testing and a native MCP server, so coding agents can pull live, current API reference and examples directly into context.

Manual prompt tuning got Dropbox a working relevance judge, then quality stalled and every model swap risked regressions. They rebuilt it as a DSPy optimization loop against an explicit objective, so the judge survives migration to smaller, cheaper models.

Dropbox explains how two engineer experiments with coding agents became a production file agent. It covers the sandboxed workspace, permission enforcement, and change recording that makes agent actions reviewable and undoable.

Dropbox scores its chat agent on the whole trajectory — intent, context gathering, tool use, grounding, completion — then closes the loop with DSPy: first calibrating the judges against human labels, then using those judges to optimize the agent's system prompt.

Dropbox engineers describe how AI-assisted development let them tackle complex integrations before launch, shifting from per-step assistance to delegated tasks, with broader engineer ownership and quality checks moved earlier in the process.

Dropbox argues the productivity story has moved past code generation: once AI raised coding throughput, the pressure landed on review queues, CI, validation, and release coordination. What it takes to go from copilots to agents executing scoped tasks.

Dropbox built Nova, an internal service that runs coding agents in its cloud — parallel sessions for engineers plus an API other internal systems call to put agents inside automated workflows. Covers why a platform beat point solutions, and what broke along the way.

Dash retrieval is only as good as the relevance labels its ranker trains on, and human judgments do not scale to billions of enterprise documents. Dropbox describes using LLMs to amplify that human labeling and lift ranking quality where it constrains answers.

Dropbox engineers describe evolving the Reclaim calendar assistant to accept natural-language requests. The design connects conversational goals to calendar context and preferences while keeping the existing scheduling engine.

Dropbox describes what changed after rolling AI out company-wide: bottlenecks shift as output grows, workflows and incentives need re-examining, and human judgment on problem selection and evaluation becomes more critical, not less.

The content-transformation platform behind Dropbox previews, Sign, and Replay has become the extraction layer feeding AI products like Dash — and is now being opened to developers through API and MCP tools. A look at how that plumbing is built to scale.