Case study · 2026
Multi-agent PR review pipeline
Built it; adopted team-wide
A tiered pipeline of Claude agents that reviews pull requests: triage, specialist reviewers, and an adversarial verifier that challenges every finding before it is posted. Part of the Claude Code + MCP toolchain I built for the team.
The problem
Code review was a bottleneck measured in hours, and a single general-purpose LLM reviewer is both expensive and noisy: it spends tokens everywhere and posts findings nobody asked for.
The goal was reviews that are cheaper, faster, and trustworthy enough that engineers read them.
One pull request through the pipeline
- pull requestdiff and metadata come in
- triage agentdecides what the change needs reviewed
- specialist agentsfocused reviewers work only on what triage flagged+ MCP code searchcontext from 9 repos, not just the diff
- adversarial verifierchallenges each finding; weak ones are dropped
- review postedonly verified findings reach the PR
Key decisions
Tier the agents
A triage pass decides which specialists run, so the expensive review work only happens where the change needs it. This is where the token savings come from.
Verify before posting
A separate agent argues against each finding. Findings that don't survive are never posted, which keeps the signal high enough that people actually read the review.
Give reviewers real context
Specialists query the custom MCP code-search servers, so they can check how code is used across 9 repositories instead of judging the diff in isolation.