Asana Accelerates Legacy Code Migration Using OpenAI Codex

Asana has successfully utilized OpenAI Codex to migrate away from Enzyme, an outdated testing system, completing a project in approximately two weeks that was previously estimated to take five years and $6 million. This demonstrates the ability of AI agents to handle large-scale codebase changes, allowing engineers to focus on higher-level craft and making previously impractical software migrations practical.

Technical Implementation of the Enzyme Removal

Asana's engineering team used Codex, powered by frontier models, to automate the removal of Enzyme. The process involved deploying up to four coding agents working in parallel across separate copies of the codebase. These agents were guided by a a five-sentence prompt, and the migration was completed in 1.5 weeks of engineering effort spread over two calendar weeks.

Key technical details of the process include:

• Parallel Agent Execution: Up to four agents operated simultaneously on separate codebase copies to accelerate the work.
• Prompting Strategy: The team found that simpler instructions worked more effectively than elaborate setups.
• Human-in-the-Loop Review: An engineer checked progress twice daily and reviewed every single proposed change to ensure quality and accuracy.

Cost and Efficiency Gains

The migration from Enzyme to a modern frontend stack was a significant cost reduction. Asana reported that the model and infrastructure costs for the AI-driven approach totaled approximately $12,000, compared to the previous staffing plan estimate of $6 million.

Implications for Software Engineering

The success of this migration has changed Asana's perspective on long-running software projects. According to Amritansh Raghav, Chief Technology Officer of Asana, "Not every years-long project will collapse into weeks. But agents can give engineers more room for craft—and the work that was once impossible now becomes worth attempting."

With the Enzyme migration complete, Asana intends to apply this agent-based approach to other migrations, rewrites, and performance problems that were previously assumed to take years to resolve.