OpenAI Codex cuts Asana 5-year testing migration to 2 weeks
Asana's engineering organization used OpenAI Codex to remove Enzyme, an old testing tool that had fallen out of active maintenance and was blocking modernization of the company's frontend stack. The migration was originally estimated to need at least five years of staffing and roughly $6 million. Using Codex, Asana finished it in about two calendar weeks, with 1.5 weeks of actual engineering effort, at a model and infrastructure cost of about $12,000.
The setup was simple: engineers wrote a five-sentence prompt, and up to four coding agents ran in parallel, each working in its own separate copy of the codebase. An engineer checked on progress twice a day and reviewed every change the agents proposed before it landed. According to the source, simpler instructions worked better than a more elaborate setup for directing the agents.
Asana already uses AI agents and automation inside its own product to help customers manage and track work, and applies the same approach internally: engineers use Codex, powered by frontier models, to tackle large codebase changes, then review and approve each one. The Enzyme removal is presented as a case study of that internal practice. With the migration done, Asana says it can now test agents on other migrations, rewrites, and performance problems it had previously assumed would take years, and hopes the approach frees engineers to spend more time on craft while people keep reviewing the work.
Key facts
- Asana used OpenAI Codex to remove Enzyme, an outdated testing system, from its codebase.
- The work was done in about two calendar weeks (1.5 weeks of engineering effort), versus a prior staffing estimate of at least five years.
- Model and infrastructure costs were about $12,000, compared with an estimated $6 million for the original staffing plan.
- Up to four coding agents worked in parallel from a five-sentence prompt, each in a separate copy of the codebase, with an engineer checking progress twice a day and reviewing every change.
- Asana says it can now consider using agents for other migrations, rewrites, and performance work it previously assumed would take years.
Why it matters
The case study is offered as evidence that AI coding agents can make large, previously impractical engineering projects, the kind of multi-year legacy migration companies often defer indefinitely, actually worth attempting. The gap between the two estimates, five years and roughly $6 million versus two weeks and about $12,000, is the headline claim: a project that could not clear a normal budget or timeline was small enough to just do.
Who it affects
Engineering leaders and teams carrying old, hard-to-upgrade code, particularly outdated testing or build tooling that blocks broader modernization. It also matters to anyone evaluating whether AI coding agents like Codex are ready for large-scale, multi-file refactoring rather than small, isolated changes.
How to use it
Asana's process: write a short prompt (five sentences in this case), run up to four coding agents in parallel, each operating on its own separate copy of the codebase, and have an engineer check in twice a day and review every proposed change before it merges. The source notes that simpler instructions outperformed a more elaborate setup. No pricing or licensing terms for Codex itself are given beyond the roughly $12,000 total cost of this specific project.
How solid is it
This is a single case study published directly by OpenAI, promoting its own product, so it is vendor content rather than independent verification. It names no individual engineer, executive, or reviewer, gives no date for when the project ran, and does not break down how the $12,000 splits between model usage and infrastructure. The $6 million, five-year figure is described as Asana's own prior estimate for a staffing plan that never happened, not a cost that was actually incurred.
Risks and caveats
The source itself cautions that 'not every years-long project will collapse into weeks,' framing Enzyme's removal as a favorable case rather than a general result. Humans still checked in twice daily and reviewed every change, so the process was not unsupervised. As a one-company, vendor-published example with no named sources and no breakdown of costs or methodology, it should be read as a promotional case study rather than a benchmark.
“Not every years-long project will collapse into weeks. But agents can give engineers more room for craft, and make once-impossible work worth attempting.”
— unnamed speaker, quoted in OpenAI's case study on Asana