Scaling Engineering Through Parallel AI Agents

Asana successfully removed the deprecated Enzyme testing framework from its codebase by deploying AI agents powered by OpenAI Codex. Instead of manual refactoring, the team utilized a parallelized agent architecture where up to four agents worked simultaneously in isolated copies of the codebase. This approach effectively turned a massive, multi-year technical debt project into a manageable, automated workflow.

The Efficiency of Simple Instructions

Contrary to the instinct to over-engineer prompts, Asana found that simpler instructions yielded superior results. The process relied on a concise, five-sentence prompt to guide the agents. Human oversight was maintained through a high-leverage review process: engineers checked the agents' progress twice daily and performed a final review and approval of every proposed code change. This human-in-the-loop model ensured code quality while offloading the repetitive, time-consuming migration tasks to the AI.

Shifting the Threshold for Technical Debt

The project's success has fundamentally changed Asana's internal perspective on what constitutes a "practical" software project. By reducing the cost of this migration from an estimated $6 million in staffing to approximately $12,000 in model and infrastructure costs, the company has lowered the barrier for tackling other long-running migrations, rewrites, and performance optimizations. This shift allows engineering teams to focus on higher-level product craft rather than being bogged down by legacy maintenance.