Asana Replaced 5-Year Migration in 2 Weeks for $12K Using Codex

Asana's engineering team removed Enzyme — a deprecated frontend testing library blocking stack modernisation — across 1.5 weeks of engineering effort spread over two calendar weeks, at a model and infrastructure cost of roughly $12,000. The previous staffing plan for the same work carried an estimated price tag of approximately $6 million and a projected timeline of at least five years. That cost compression and the collapse from a multi-year roadmap item into a two-week sprint is the figure engineering leaders evaluating AI coding agents should anchor on.

Why This Migration Was the Test Case

Enzyme had fallen out of active maintenance, and its presence in Asana's frontend codebase was actively blocking dependency upgrades and stack modernisation. The migration was prerequisite work, not discretionary cleanup — its long projected duration had frozen a chunk of Asana's upgrade roadmap indefinitely. Choosing this category of task — mandatory, unglamorous, high-blast-radius — as the Codex test case reflects a framing of agent utility centred on making previously impractical remediation work worth attempting, not accelerating greenfield development.

Deployment Architecture

Asana generated the entire migration from a five-sentence prompt. Up to four Codex agents ran in parallel, each operating inside a separate copy of the codebase rather than a shared working tree. That isolation matters: parallel agents writing to a single repo produce merge conflicts and state entanglement; sandboxed copies allow independent task execution and cleaner diffs at review time. An engineer reviewed every proposed change and checked agent progress twice daily — human review was continuous but lightweight, not a bottleneck. This aligns with patterns in multi-agent system design where control layers remain human-owned even as execution parallelises.

Asana also found that simpler prompting outperformed a more elaborate setup — a recurring empirical finding in production agentic deployments worth weighing against the instinct to over-engineer task specifications.

The Economics

The $12,000 figure covers model API usage and infrastructure only — it does not represent zero human labour. Engineer time for twice-daily progress checks and change review over two calendar weeks is a real cost, but it is a rounding error relative to the $6 million staffing baseline. The more important structural shift is from headcount — a recurring, capacity-constrained resource — to inference spend, a variable and instantly scalable one. Teams modelling the economics of AI coding agents should capture that distinction, not just the per-task dollar figure.

The security surface of agentic coding deployments also warrants explicit attention in any evaluation. When agents operate across large codebases with write access, the trust and verification architecture matters — a concern directly relevant to how coding agents interact with production systems and credentials.

What Comes Next

With Enzyme removed, Asana's stated intent is to apply the same model to other migrations, rewrites, and performance problems previously categorised as impractical given timeline and cost. CTO Amritansh Raghav was explicit that not every years-long project will compress this dramatically — an important qualifier. Enzyme migration is a well-scoped, largely mechanical transformation with clear success criteria, making it closer to an ideal agent task than ambiguous feature development.

The Asana case establishes a concrete enterprise benchmark: a real production codebase, a mandatory migration, parallel agents with human review, two weeks, $12,000. Engineering teams evaluating Codex or comparable agents now have a reference point that is not synthetic. Whether the pattern generalises to less mechanically bounded work is the open question the industry will spend the coming year answering.