Sidekick is HubSpot's internal AI code reviewer, built by its Developer Experience AI team. It runs on every pull request in GitHub. A second Judge Agent evaluates each draft comment for succinctness, accuracy, and actionability, and only comments that pass are posted.
Pull request opened → Sidekick draft review → judge-gated comments posted to GitHub.
Trigger review
Review runs automatically when a pull request is created or marked ready for review, or when Sidekick is requested as a reviewer[Source 1, Source 2]
Draft the review
Sidekick produces the initial review of the pull request[Source 1]
Post passing comments
Only comments that pass the judge evaluation are posted to GitHub; the rest are filtered out[Source 1]
Where people stay involved
People review, approve, or guide the agent’s work here.
Not reported: The review flow is automated end to end. The captures document post-hoc emoji and thread feedback on comments (source 1, lines 103–108) and continued human review for system design and architecture (line 12), but no required human checkpoint inside the review flow.
Not reportedNo capture states which model the review runs on. Multi-model support (Claude, GPT, Gemini) is a capability of the Aviator framework (source 1, line 45), not a statement about this reviewer.
Harness
Aviator, an internal Java agent framework; replaced the earlier Claude Code review implementation on Crucible[Source 1]
Sandbox
Not reportedExecution isolation of the Aviator runtime is not documented. The Kubernetes sandboxing described in source 2 (lines 30–37) covers the earlier Crucible implementation.
Tool access
Aviator framework for precise tool control[Source 1]
Knowledge
Optional repo-specific custom instructions can be incorporated into the review prompt; the feature recently launched[Source 1]
Context management
Not reportedThe captures describe no mechanism that maintains or organizes context for the reviewer over time.
A second Judge Agent evaluates every draft comment for succinctness, accuracy, and actionability[Source 1]
Reported observations
These results come from the sources. Notes below each result explain what it covers and any reported limits.
Headline claim
Reviews every pull request and cut engineer feedback time by 90%[Source 1]
Reported by
HubSpot
Scope
Time for engineers to receive code feedback from Sidekick; not overall PR completion time
The source does not report the denominator of this figure.
Observed in March 2026
About this result
Category
Cost latency
Evidence type
Reported measurement
What it covers
Sidekick review coverage of HubSpot pull requests and the time engineers wait for feedback
Key observation
Over 80% thumbs-up reaction rate on review feedback during the preceding couple of months[Source 1]
Reported by
HubSpot
Scope
Developer emoji reactions on review comments during the preceding couple of months
Denominator
Thumbs-up and thumbs-down reactions; not all developers or all reviews
Method
Emoji reactions and replies on review comments
About this result
Category
Effectiveness
Evidence type
Reported measurement
What it covers
Developer thumbs-up share of reactions on Sidekick review comments over the preceding couple of months
Lessons
Reported opinion: Unhelpful rather than incorrect feedback was the most common failure mode; prompt tuning alone could not eliminate it, and a second judge agent filtering draft comments was the single most important factor in review effectiveness[Source 1]
Not reported: The review flow is automated end to end. The captures document post-hoc emoji and thread feedback on comments (source 1, lines 103–108) and continued human review for system design and architecture (line 12), but no required human checkpoint inside the review flow.
Duplicate representation of Headline claim. Restates the every-pull-request coverage component of the compound headline; same passage, scope, and period.
Statement type
Metric
Provenance
Reported
Confidence
Medium
Confidence reason
HubSpot reported the figures in its own engineering blog.
