Nova is Dropbox's internal service for running coding agents in its cloud. Engineers launch parallel sessions from a web UI, CLI, or API, and internal systems call agents inside automated workflows like CI triage and migrations. Callers attach validation commands that Nova runs after each attempt.
Multiple coding agents behind one interface; prompt-evaluation tooling; helpers that make it easier to add AI-powered steps without rebuilding surrounding infrastructure; sources name no specific model or provider[Source 1]Multiple coding agents run behind one interface; no specific model or provider is named.
Harness
Validation loop (propose → validate → feed back) with continue_on_validation_failure and max_iterations (~5); branch management kept outside the agent[Source 1]
Sandbox
Isolated env with a codebase snapshot at a specific commit; full Dropbox monorepo via Bazel; hermetic remote execution + caching[Source 1]
Localized AGENTS.md per service; Dash (Dropbox context engineering); passing + failing test logs[Source 1]
Context management
Session history (notes/logs) carried across retry attempts[Source 1]
Credentials
Not reportedThe capture states that agents operate within existing infrastructure and validation paths but names no authentication or authorization mechanism.
Bounded iteration with feedback on failure; deterministic systems control test execution[Source 1]
Dash
Nova is expanding its context sources, including Dash and MCP-based integrations[Source 1]
Documented uses
Athena flaky-test alert through the Deflaker fix-and-validate loop to a landed fix or capped attempts.
Ingest flaky-test evidence
Athena detects a flaky test; the Deflaker workflow sends its passing and failing logs to Nova as context and asks the agent to identify a likely root cause[Source 1]
Propose a root-cause fix
The Nova agent proposes a fix for the flaky test[Source 1]
Retry with carried-forward notes
A flake starts another attempt with new logs and notes from the previous one, until a working fix lands or attempts reach the cap of five[Source 1]
Access and controls
No separate access-control assessment is recorded.
Reliability and validation
Validate the fix in CI
CI runs the test 100 or more times depending on its failure rate to check the proposed change[Source 1]
Adoption and operating evidence
These results come from the sources. Notes below each result explain what it covers and any reported limits.
Headline claim
Dozens of agents can run in parallel from one runbook[Source 1]
Reported by
Dropbox
Scope
Migration-owner orchestration of dozens of agents from a shared runbook; qualitative capacity description
About this result
Category
Runtime capacity
Evidence type
Qualitative
What it covers
Parallel agents a migration owner launches and manages from one runbook
Key observation
Roughly 1 in 12 pull requests at Dropbox, with adoption continuing to grow, as of May 2026[Source 3]
Reported by
Dropbox
Scope
Share of all Dropbox pull requests accounted for by Nova, snapshot at publication
Reported: The Deflaker run is the documented operational workflow; interactive sessions and the crash-alert experiment are other modes on the platform, outside this scope.
human involvement
Not applicable: There is no single platform-wide agent review boundary. Scoped downstream operating-model claims remain in research details; access controls are described separately.
implementation
Reported
validation
Reported: Caller-attached validation commands run after each attempt; the Deflaker CI runs are the documented instance.
Duplicate representation of Headline claim. Same dozens-of-agents capacity from the same runbook passage; identical scope with no separate period or qualification.
Statement type
Metric
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
Reported by
Dropbox
Scope
Migration-owner orchestration of dozens of agents from a shared runbook; qualitative capacity description
The retrospective explicitly makes this assessment and names context files, tests, caching, isolation, and retry loops; it provides no controlled component comparison.
Unclassified for event-triggered validation-gated remediation (Deflaker, crash-alert candidate fixes) → landed fix or candidates routed to service teams; human attention boundary: unknown.
Statement type
Inference
Provenance
Catalog judgment
Confidence
Unverified
Confidence reason
Deflaker automates bounded repair and CI validation, while publication stays outside the agent. The source does not specify a required human approval or exception-handling boundary.
Level 2 for interactive developer session from web, CLI, or Slack → agent-assisted change; human attention boundary: continuous steering.
Statement type
Inference
Provenance
Catalog judgment
Confidence
Medium
Confidence reason
The source describes standard interactive chat sessions in which engineers launch and steer agents; the exact review boundary of those sessions is not further specified.
