Find and verify bugs before opening a PR
A tool-using agent loop runs a finder pass to list bugs in code or prose, then a verifier pass tries to prove or disprove them before returning results. ·
Done with a custom-built or unnamed agent
- What they did
- The tool, called roast, is an agentic loop that uses tools and makes two separate passes over code or prose before a PR is opened. A finder pass looks for bugs and outputs a structured list of issues. A verifier pass then tries to prove or disprove each item. Only then are results returned to a human or to a harness.
- How it went
- The post is a short announcement. It gives no results, numbers, or examples of bugs caught, and it says nothing about false-positive rates.
Try it yourself with Claude Code
In my repo at [path or repo name], review the changes in [branch or PR] and run two passes. First, list every possible bug you can find. Second, act as a skeptical verifier: for each suspected bug, try to prove or disprove it by reading the code and running tests or small scripts. Give me a final list of only the confirmed bugs with file, line, evidence, and a suggested fix, and don't change any files until I approve.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: Bluesky · Nov 5, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Build a searchable map of funded companies
A software engineer had Replit Agent build a searchable map of Wefunder brick-and-mortar companies in about 3 hours for about $22.
Review GitHub pull requests automatically
An n8n AI agent triggers on each new PR, fetches diffs, calls an LLM for a review, and posts comments and labels back on the PR.
Build a multi-tenant SaaS with an agent
The poster built a production multi-tenant SaaS in 15 days with Replit Agent, but hit issues like RLS bypasses and production pointing at the dev database, so checking its work was essential.
Plan with PRDs and delegate tasks to cheaper agents
A developer uses a high-level agent to write a PRD, splits it into a task list, and hands the work to cheaper agents in a production workflow.
Review agent work plans and track user intent logs
A developer logs every user message to distill intent and has agents review saved plan files against project constraints before executing work.
Custom agent
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.