How to use Codex to audit a large codebase overnight
The author ran approximately 800 narrowly scoped auditors overnight, assigning one file per subagent. Later synthesis extracted bugs, identified common causes and grouped findings by priority.
news.ycombinator.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Audit [repository] without changing source files. Assign one file per worker and run at most [batch size] workers concurrently. Write evidence-backed reports to [directory], then synthesize bugs, shared causes and priorities without treating unverified findings as confirmed.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- esperent
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reports approximately 800 overnight auditors producing reports for later synthesis. A separate restricted repair run cleared a few thousand lint and type issues over several hours.
- Human involvement
- The author defines worker scope and runs subsequent synthesis; nontrivial repair findings are reserved in a TODO file for human review.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Assign one source file and a report-writing task to each subagent.
- Launch auditors in batches of 50 through the coordinating session.
- Collect the recommendations written by the overnight auditors.
The author describes running read-only audits of very large codebases overnight. Each subagent receives one file and writes a report with recommendations. The work runs through pi and a subagent extension that accepts shared prefix and suffix prompts alongside a JSON list of names and specific prompts.
The coordinating session is instructed to launch workers in batches of 50 because the author encounters Codex rate limits at roughly that concurrency. Individual auditors reportedly used between 7,000 and 40,000 tokens of context, and one overnight run completed approximately 800 small audits.
After the workers finish, the author runs synthesis over their written reports. That pass extracts bugs, followed by another round that identifies common sources and groups findings by priority. The reports therefore become inputs to a separate consolidation process.
The author also describes a related repair experiment: waves of workers fixed only obvious lint and type issues, while other findings went into a TODO file for human review. They report clearing a few thousand issues over several hours.
Who it's for: Maintainers of large repositories who can divide audits into narrow assignments and review consolidated findings.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · esperent
Frequently asked questions
Why did the author launch auditors in batches of 50?
The author reports encountering Codex rate limits at about 50 concurrent subagents, so the coordinating session was instructed to use batches of that size.
How were hundreds of audit reports consolidated?
The author ran synthesis over the written reports to extract bugs, then used another round to find common sources and group findings by priority.
Were workers allowed to fix every issue they found?
In the separate repair experiment, workers were instructed to fix only obvious or trivial lint and type issues. Other issues were appended to a TODO file for the author's review.
More details
Remaining steps
- Synthesize the reports to extract individual bugs.
- Run another consolidation pass to identify shared causes and priorities.
About this account
A coordinating session dispatches narrowly scoped auditors in batches, allowing read-only analysis to run overnight.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=48642750. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: esperent on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
Codex · pi · Subagent extension · JSON task lists
Code audits · Multi-agent · Large codebases · Bug triage
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.