---
type: "post"
slug: "agent-builder-ops-adds-review-mocks-cost-controls-and-secure-workflows"
title: "PR-AF turns code review into a planned multi-agent pass"
description: "PR-AF builds a per-PR review plan, spawns focused reviewer agents, grounds findings in code evidence and challenges them — reporting 0.706 golden recall as the top open-source reviewer."
retrieval_nugget: "PR-AF builds a per-PR review plan, spawns focused reviewer agents, grounds findings in code evidence and challenges them — reporting 0.706 golden recall as the top open-source reviewer."
published_at: "2026-08-14"
updated_at: "2026-09-12"
record_date: "2026-08-14"
date_kind: "discovered_at"
topics: ["coding-agents","evals","security"]
entities: ["PR-AF","aimock","LangSmith","GitHub"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:agent-builder-ops-adds-review-mocks-cost-controls-and-secure-workflows"
status: "published"
source_urls: ["https://github.com/Agent-Field/pr-af"]
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:31aac59f304eeaf4c7b41b83043f0155ef5aa8c1e397d896ce53a5e61185229f","reviewed_at":"2026-09-12T10:00:00Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":1,"watch_condition_count":1,"related_record_count":3}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"PR-AF is an open-source code reviewer built around a pipeline rather than a prompt: turn each pull request into a task-specific review plan, spawn focused reviewer agents, ground findings in code evidence, then challenge the results.","observed_facts":[{"text":"PR-AF reports 0.706 golden recall, first among 42 compared tools, and 595 independently valid findings.","source_urls":["https://github.com/Agent-Field/pr-af"]},{"text":"It publishes public results, per-pull-request judge verdicts and reproduction scripts, and runs locally in Docker.","source_urls":["https://github.com/Agent-Field/pr-af"]}],"mechanism":"it maps model class to review stakes — DeepSeek-class models for routine pull requests, GLM-5.2 for deep open-model reviews, frontier Opus-class models for major ones","why_now":"The reported numbers are unusually specific for this category: 0.706 golden recall, described as first among 42 compared open-source and commercial tools, and 595 independently valid findings, roughly three times the leading commercial tools in an adjusted comparison.","implications":["Challenging your own findings is the part that matters, because an unchallenged reviewer optimises for volume and a human then pays for every false positive."],"evidence_boundary":"The benchmark is Code-Review-Bench and the results are self-reported by the tool's authors, with the adjusted comparison defined by them.","watch_conditions":["The condition to watch is whether an independent run reproduces the recall figure on unseen pull requests, because a reviewer evaluated on a public benchmark faces exactly the optimisation problem its own challenge step is meant to catch."],"related_records":[{"url":"https://newruntime.com/patterns/verification-bandwidth-is-the-scarce-resource","relation":"Existing pattern on review capacity as the constraint."},{"url":"https://newruntime.com/posts/a-three-person-team-shipping-hundreds-of-prs-moves-the-bottleneck-to-review","relation":"Prior coverage of the review bottleneck in practice."},{"url":"https://newruntime.com/topics/coding-agents","relation":"Topic hub for the code review layer."}]}
routes: {"html":"https://newruntime.com/posts/agent-builder-ops-adds-review-mocks-cost-controls-and-secure-workflows/","markdown":"https://newruntime.com/posts/agent-builder-ops-adds-review-mocks-cost-controls-and-secure-workflows.md","json":"https://newruntime.com/posts/agent-builder-ops-adds-review-mocks-cost-controls-and-secure-workflows.json"}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
---

# PR-AF turns code review into a planned multi-agent pass

## Retrieval answer

PR-AF builds a per-PR review plan, spawns focused reviewer agents, grounds findings in code evidence and challenges them — reporting 0.706 golden recall as the top open-source reviewer.

PR-AF is an open-source code reviewer built around a pipeline rather than a prompt: turn each pull request into a task-specific review plan, spawn focused reviewer agents, ground findings in code evidence, then challenge the results. Its stated aim is to squeeze more useful review out of cheaper models, and it maps model class to review stakes — DeepSeek-class models for routine pull requests, GLM-5.2 for deep open-model reviews, frontier Opus-class models for major ones.

The reported numbers are unusually specific for this category: 0.706 golden recall, described as first among 42 compared open-source and commercial tools, and 595 independently valid findings, roughly three times the leading commercial tools in an adjusted comparison. It publishes public results, per-pull-request judge verdicts and reproduction scripts, runs locally in Docker, and can be triggered by CLI, curl, CI or another agent.

Challenging your own findings is the part that matters, because an unchallenged reviewer optimises for volume and a human then pays for every false positive. That is the constraint recorded in [verification bandwidth being the scarce resource](https://newruntime.com/patterns/verification-bandwidth-is-the-scarce-resource/) and observed where [a three-person team shipping hundreds of pull requests moved the bottleneck into review](https://newruntime.com/posts/a-three-person-team-shipping-hundreds-of-prs-moves-the-bottleneck-to-review/). The [coding agents hub](https://newruntime.com/topics/coding-agents/) tracks the review layer.

The benchmark is Code-Review-Bench and the results are self-reported by the tool's authors, with the adjusted comparison defined by them. The condition to watch is whether an independent run reproduces the recall figure on unseen pull requests, because a reviewer evaluated on a public benchmark faces exactly the optimisation problem its own challenge step is meant to catch.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- topic hub: [Evals](https://newruntime.com/topics/evals.md): Explore the evals topic hub.
- topic hub: [Security](https://newruntime.com/topics/security.md): Explore the security topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)