Asana's 5-year frontend migration took 2 weeks with Codex. The workflow is the story. (2026)

1/10

Asana says a frontend migration it expected to take five years took about two weeks with Codex.

The reported bill for models and infrastructure: about $12,000, versus a roughly $6 million staffing estimate for the old plan.

The interesting part is not the ratio. It is the workflow: a short prompt, up to four agents working in parallel, separate codebases, twice-daily checks, and human review of every proposed change.

This is a close read of OpenAI's August 18, 2026 case study, Asana cleared 5 years of engineering work in 2 weeks with Codex. 1

2/10 — The job was painfully specific

Asana wanted to remove Enzyme, an older testing system from its frontend stack.

The source says Enzyme had fallen out of active maintenance and was making future frontend upgrades harder.

That matters because the agents were not asked to "improve the codebase" in the abstract. They had a named dependency, a defined end state, and a large but bounded migration to complete. 1

A useful AI-coding task starts with that level of specificity.

3/10 — The setup was smaller than the hype

Asana gave Codex a five-sentence prompt.

Up to four coding agents then worked in parallel, each inside a separate copy of the codebase.

The separate copies created room for parallel exploration without putting several agents directly into the same working tree.

This was a migration workflow, not one agent improvising across the production repository. 1

4/10 — The human gate stayed in the loop

An engineer checked progress twice a day and reviewed every proposed change.

Asana's description says engineers used Codex for large codebase changes, then reviewed and approved the proposed work.

That changes what the result means. The agents accelerated search, edits, and iteration. A person still decided whether the migration was correct enough to keep. 1

The reusable pattern is autonomy between checkpoints, not autonomy without checkpoints.

5/10 — The calendar time hides a useful detail

The source says the migration took 1.5 weeks of engineering effort spread across two calendar weeks.

Enzyme was fully removed at the end of that period.

That is different from saying four agents ran unattended for two weeks. Engineers supplied the initial task, monitored progress, and reviewed the output along the way. 1

When teams report an agent result, effort time and elapsed time should be separate fields.

6/10 — Treat the cost headline carefully

OpenAI's case study puts model and infrastructure costs at about $12,000.

It compares that figure with Asana's earlier plan: at least five years of work and roughly $6 million in staffing costs.

That is a useful scale signal, but it is not a like-for-like budget. One number covers models and infrastructure; the other is a staffing estimate. The article presents both as Asana's comparison. 1

The honest takeaway is that the agent workflow made the project economical enough to attempt.

7/10 — Simpler instructions won

One of the most practical lines in the case study is easy to skip: simpler instructions worked better than a more elaborate setup.

That is a useful correction to the instinct to build a giant prompt, a complex agent hierarchy, and a long rulebook before testing the task.

Start with the smallest instruction that defines the migration, the constraints, and the finish line. Add machinery only when the work shows you where it is needed. 1

8/10 — What can builders copy?

The source says the experience changed which long-running migrations, rewrites, and performance problems Asana considers practical.

A compact version of that pattern looks like this:

Choose a codebase change with a clear end state.
Give each agent an isolated copy.
Use a short first instruction.
Check progress on a fixed schedule.
Review every proposed change.
Measure elapsed time, human effort, and model cost separately.

The workflow works best when agents can explore in parallel while the team keeps one accountable approval path. 1

9/10 — The quote worth keeping

Asana CTO Amritansh Raghav puts the claim in the right register:

"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."
—Amritansh Raghav, Chief Technology Officer, Asana 1

The second sentence is the durable idea. The value is not that every migration becomes a two-week task. The value is that some work crosses the threshold from "too expensive to start" to "worth a serious attempt."

10/10 — The real lesson is the gate, not the agent count

If you want to test this workflow, start with one migration that has:

an explicit finish line;
a codebase that can be copied or sandboxed;
tests that can catch regressions;
scheduled human review; and
a budget that counts model, infrastructure, and engineering time separately.

Asana's case study is a company-reported example, not a promise that every five-year project will collapse into two weeks.

But it does offer a concrete builder brief: use agents to widen the search space, keep the work reviewable, and let humans own the merge.

What migration in your codebase is still waiting for the economics to change?

Fuentes de referencia

1
Asana cleared 5 years of engineering work in 2 weeks with Codex
openai.com

Fuentes de referencia

Contenido relacionado