Skip to content

Let a coding agent write most new features and fix defects 10x faster

Cisco rolled out OpenAI Codex across engineering and reports it wrote over 95% of the new features in its AI Defense product. Defect-fix throughput rose 10-15x, and the company says it saves more than 1,500 engineering hours a month.

OpenAI Codex One public report
52
Proof
18/40
Payoff
25/25
Ease
7/20
Cost
2/15
How the score works
Difficulty
Builder
Setup time
1 hr
Monthly cost
$200/mo
What it saves
1,500+ eng hours per month

How to do it

  1. Start with a backlog of well-described bugs from your static analysis tools
  2. Run the Codex CLI on each defect with the repo and the failing check
  3. Have engineers review and merge the fixes
  4. Expand to cross-repo work, such as build speedups and framework migrations
  5. Track hours saved and defects closed per week

The prompt

Here is a static analysis finding: <paste finding>. Find the root cause in this C++ codebase, not just the reported line. Propose the smallest safe fix, apply it, build the project, and run the related unit tests. Summarize the cause, the fix and any other places with the same pattern.

Where this comes from

One person reported it. Promising, not yet confirmed. Original report: OpenAI.

Tried this? Tell us how it went and help move its proof level up.

Was this useful?

More like this

  • 67

    Move 2,500 marketing pages off a CMS and cut build times from 14 to 4 minutes

    Sentry's marketing and growth team, with a couple of developers, used Claude Code to move about 2,500 pages from a headless CMS into Markdown on Astro, folding 200 one-off pages into 3 templates. Builds dropped from about 14 minutes to under 4, and non-engineers now ship pages through Claude skills.

    Claude Code Several reports Builder $100/mo
  • 67

    Have an agent code a complex open-source feature unattended for seven hours

    Rakuten ML engineer Kenta Naruse gave Claude Code the job of adding an activation vector extraction method to the vLLM library. It worked on its own for seven hours, he wrote no code, and the result matched the reference method with 99.9% numerical accuracy.

    Claude Code Several reports Builder 1 hr setup $100/mo
  • 66

    Build a newsletter summarizer that cuts weekly reading to 30 minutes

    Consultant Justin Tan, a self-described low-coder, used Replit Agent over three days to build an app that pulls newsletters from Gmail, removes duplicate stories, summarizes them and archives the emails. His weekly newsletter reading dropped from 2-3 hours to two 15-minute sessions.

    Replit Agent One report Beginner 4 hr setup $25/mo
  • 66

    Replace a $40K CRM contract with a custom CRM built in an afternoon

    Atonom, a 25-30 person AI sales company, built its own CRM in Lovable with a first working prototype in about 3 hours. It replaced a Salesforce contract, cutting annual CRM costs from $40,000 to about $1,200.

    Lovable One report Beginner 3 hr setup $25/mo