Asana, the company behind the work management platform, says it cleared a code cleanup it had scoped at five years in just two weeks using OpenAI Codex[1]. Model and infrastructure costs came to about 12,000 USD (about 1.9 million yen), against roughly 6 million USD (about 950 million yen) for the staffing plan it had drawn up. Here is what the company handed over to the agents, and how.
Five years of scope, compressed into two weeks
The target was Enzyme, an outdated testing tool that Asana wanted out of its codebase[1]. The frontend depended on it, and that dependency had become a blocker for modernizing the underlying stack.
Engineers spent about 1.5 weeks of actual effort, spread across two calendar weeks[1]. The earlier plan called for at least five years and an estimated 6 million USD. Model and infrastructure costs for the Codex run totaled roughly 12,000 USD[1].
※1 USD = 159 JPY (as of August 18, 2026)
A five-sentence prompt and up to four agents in parallel
The setup was deliberately plain. Everything started from a five-sentence prompt, and up to four coding agents ran in parallel from there[1]. Each agent worked inside its own separate copy of the codebase, so their changes never collided.
Human involvement stayed narrow. One engineer checked progress twice a day and reviewed every proposed change before approving it[1]. Asana's own takeaway is that simpler instructions beat a more elaborate setup[1]. Splitting the work into parallel slices mattered more than trying to control the agents with richer context.
That review-everything posture mirrors how Asana runs its own product. Its platform uses AI agents and automations to help customers manage, track, and execute work, and the company applies the same approach inside its engineering organization[1].
Why Enzyme had turned into an immovable block
Enzyme was once a widely used library for unit testing React components. It failed to keep up when React changed its internals in v17 and beyond, and no official adapter for React 18 exists, leaving the project effectively unmaintained[2].
Any project sitting on a large pile of Enzyme tests therefore stalls the moment it tries to move to a newer React. One developer who attempted the migration built a compatibility layer to translate Enzyme API calls into a modern testing library, abandoned it because the two designs differ too fundamentally, and settled on running an older React only for the test suite[2]. If that is the situation at a few hundred tests, a five-year estimate at Asana's scale is not an exaggeration.
Enzyme also encouraged tests that inspect internal implementation details, such as component state and individual methods, which made mechanical bulk replacement impractical. Each test had to be read for intent and rewritten, exactly the kind of work that quietly consumes headcount.
Work that was written off as not worth doing comes back
With the migration finished, Asana can now try agents on other migrations, rewrites, and performance problems it had assumed would take years and had never started[1].
The comment OpenAI quotes from Asana is measured about what this implies. Not every years-long project will collapse into weeks, the company says, but agents can give engineers more room for craft and make once-impossible work worth attempting[1].
There is separate data showing how far Codex use has spread. Internal figures OpenAI published in June 2026 put Codex adoption among its own employees at 97.9 percent, up from roughly 40 percent in August 2025[3]. Even non-technical departments such as Legal and Recruiting report Codex as their primary AI tool[3].
Summary
Asana removed Enzyme, a job it had scoped at five years, in two weeks using OpenAI Codex. A five-sentence prompt kicked off up to four parallel agents, while engineers limited themselves to twice-daily check-ins and reviewing every change. Costs came to about 12,000 USD (about 1.9 million yen) in model and infrastructure spend, far below the roughly 6 million USD (about 950 million yen) alternative. The finding that simple instructions outperformed an elaborate harness is the part worth borrowing for teams weighing agent adoption.
Source[1]: https://openai.com/index/asana
Source[2]: https://darekkay.com/blog/react-18-enzyme/
