Quality assurance · Production

Dropbox rethinks engineering productivity with Nova, an internal agentic coding platform

The problem

AI coding tools dramatically increased code generation throughput at Dropbox, but this shifted bottlenecks downstream into review queues, CI systems, validation workflows, and release coordination — the broader software development lifecycle could no longer absorb the volume of AI-assisted output at the same speed it was produced.

First attempt

First-wave AI copilots only accelerated implementation work without changing surrounding systems, leaving review, testing, validation, and release processes increasingly strained. Pull request throughput alone proved insufficient as a productivity metric once AI changed the volume and shape of output.

Workflow diagram · grounded in source
1
Engineer defines task
trigger
“allow engineers to describe a task in plain language and run an AI coding agent in a controlled environment”
2
Agent executes and iterates
ai_action
“An agent can take a scoped task, inspect the codebase, edit files, run tests, iterate on failures, and return an artifact for human review”
3
Result validated against guardrails
validation
“define the task, allow the agent to execute within established guardrails, validate the result, and have a human make the final judgment before any code reaches production”
4
Human final judgment
human_review
“have a human make the final judgment before any code reaches production”
5
Quality signal feedback loop
feedback_loop
“We track signals such as code review turnaround time, first-run test pass rate, defect ratio, and rework rate to understand whether increased output is holding up under real-world conditions”
Reported outcome

Dropbox built Nova, an internal coding agent platform now accounting for roughly 1 in 12 pull requests, enabling engineers to initiate more parallel work and offload repetitive execution.
A new multi-stage measurement model tracks progress from AI usage to customer value, complemented by quality signals such as code review turnaround time and defect ratio.

Reported metrics
Nova share of pull requestsroughly 1 in 12 pull requests
Parallel work capacity for engineersinitiate more parallel work, explore more options
Repetitive execution burden on engineersoffload repetitive execution that previously consumed significant time and attention
Reported stack
Nova
Source
https://dropbox.tech/culture/beyond-code-generation-rethinking-engineering-productivity-in-the-age-of-ai-agents
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Dropbox built Nova, an internal coding agent platform now accounting for roughly 1 in 12 pull requests, enabling engineers to initiate more parallel work and offload repetitive execution.

What tools did this team use?

Nova.

What results were reported?

Nova share of pull requests: roughly 1 in 12 pull requests; Parallel work capacity for engineers: initiate more parallel work, explore more options; Repetitive execution burden on engineers: offload repetitive execution that previously consumed significant time and attention (source-reported, not independently verified).

What failed first in this deployment?

First-wave AI copilots only accelerated implementation work without changing surrounding systems, leaving review, testing, validation, and release processes increasingly strained.

How is this quality assurance AI workflow structured?

Engineer defines task → Agent executes and iterates → Result validated against guardrails → Human final judgment → Quality signal feedback loop.

Related quality assurance cases