Slow, Thorough, and Top of the Code-Review Benchmark
PR-AF runs a multi-agent cognitive pipeline on every pull request to find systemic bugs that faster, shallower tools miss.

What it does PR-AF is an agentic code-review system built on the AgentField framework. It ingests a pull request, dynamically compiles a review plan tailored to that diff’s topology, and spawns ephemeral specialist agents to hunt for bugs. The output is a set of inline GitHub comments that cite file-and-line evidence, severity ratings, and suggested fixes.
The interesting bit Instead of blasting a single prompt at an LLM, PR-AF morphs its own execution graph per PR. It clusters risks across files to detect compound vulnerabilities, then runs a falsifiability gate that actively tries to kill weak findings before they reach your inbox. The README claims this keeps false positives extremely low and lets cheaper models punch above their weight.
Key highlights
- Ranks #1 in golden recall (0.706) on the Martian Code-Review-Bench among 42 compared open-source tools, and reports roughly three times as many independently valid findings as leading commercial alternatives in the adjusted comparison.
- About 10× cheaper per review than closed-source SaaS tools, according to its own benchmark analysis.
- Model-flexible: swap in lightweight models for routine PRs, GLM-5.2 for open-source CI gates, or frontier models for high-stakes changes.
- Self-hosted via Docker or Railway; runs fully offline in your own infrastructure with your own API keys.
- Execution time is ~35–50 minutes per review, so it is explicitly positioned as a CI/CD gatekeeper rather than a real-time chat companion.
Caveats
- The project is essentially a specialized AgentField node, so you need the AgentField control plane running somewhere; it is not a standalone binary.
- That 35–50 minute runtime is a feature for depth, but it rules out interactive inner-loop use.
- The benchmark claims are specific to the Martian Code-Review-Bench snapshot; your own bug taxonomy may vary.
Verdict Teams that want deep, evidence-grounded architectural audits before merging should look here. If you need instant feedback while typing, stick with faster interactive tools.
Frequently asked
- What is Agent-Field/pr-af?
- PR-AF runs a multi-agent cognitive pipeline on every pull request to find systemic bugs that faster, shallower tools miss.
- Is pr-af open source?
- Yes — Agent-Field/pr-af is an open-source project tracked on heatdrop.
- What language is pr-af written in?
- Agent-Field/pr-af is primarily written in Go.
- How popular is pr-af?
- Agent-Field/pr-af has 570 stars on GitHub.
- Where can I find pr-af?
- Agent-Field/pr-af is on GitHub at https://github.com/Agent-Field/pr-af.