Authoring
PR Author
Owns the PR end-to-end: implements, answers comments, fixes CI and conflicts.
Code Review
When agents write the code, review becomes the bottleneck.
Current AI code review tools comment on the diff and stop.
Cosmos orchestrates specialized review agents that assess risk, catch issues, and surface the decisions that require human judgment. The whole system gets sharper every PR.
cosmos / code-review
Meet Cosmos
A team of agents on every PR.
Cosmos auto-approves low-risk changes, runs line-by-line correctness analysis on the rest, and pulls humans in only for the calls that matter.
Measured on Augment’s own codebase
3×
More code shipped
67%
Reduction in median merge time
65%
Precision on issues flagged
<1.5%
Weekly revert rate threshold
Fleet of agents takes ownership of the entire code review process.
Cosmos lets you orchestrate multi-agent, multi-step review workflows across risk, correctness, judgment, and authoring. Each loop gets better with shared memory, and each loop is a Cosmos Expert.
Review lifecycleone PR, end to end
Authoring
PR Author
Owns the PR end-to-end: implements, answers comments, fixes CI and conflicts.
Routing · Auto-approve
PR Risk Analyzer
Auto-approves low-risk PRs and routes the rest by risk dimension.
Correctness
Deep Code Review
Independent, line-by-line correctness that catches most high- and medium-severity bugs.
Judgment
Pair Reviewer
Reviews code intent across architecture, security, design, and product.
Human
Makes the high-level judgment calls.
Comments posted · merged
Code Review Memory
Captures feedback · Distills per-repo knowledge · Shared by every expert
Fig 1 · Review fleet
Built for a different reader
Optimized for recall, not precision.
Other tools optimize for precision because a human reads the output. Cosmos reviews for an agent, so it goes for recall: every issue caught, not just the few worth a person’s time.
Precision
Built for a human reader
- Surfaces only the highest-importance issues
- Lets real bugs slip through to stay quiet
- Capped at what a person will read
Recall
Built for an agent reader
- Catches every issue it can find
- Exhaustive, no triage for brevity
- Scales past what a person could read
Measured on our own codebase.
Since Cosmos took over reviews on the Augment codebase, output is up, merge times are down, and the revert rate has stayed under threshold every week.
Code output vs. merge timePRs/week + median merge minutes, weekly
Bugs per output unitBug-introducing commits per output unit, weekly
Weekly revert rateShare of merged PRs subsequently reverted
GitHub
GitLabSAML / OIDC / SCIMGDPR · CCPA · HIPAAHighly customizable to your organization.
Talk to Cosmos Advisor to tailor every agent: review policy, risk dimensions, escalation rules. Run on the VCS, identity stack, and compliance posture you already use.
GitHub or GitLab. SAML, OIDC, SCIM. GDPR · CCPA · HIPAA. BAA available. Audit logs and SIEM out of the box.