How do you estimate software projects in the age of AI agents? I gave ChatGPT the prompt in the screenshot. It estimated 18 months and $2 million for the migration scenario I described. That prompt is basically describing Asana’s Enzyme migration. Asana had estimated the migration would take at least 5 years and roughly $6 million. Using Codex, they finished it in about 𝟐 𝐰𝐞𝐞𝐤𝐬 𝐟𝐨𝐫 𝐚𝐫𝐨𝐮𝐧𝐝 $𝟏𝟐,𝟎𝟎𝟎. Bun’s recent migration is another example of agents doing work that would normally take years much faster. Jarred Sumner estimated that porting Bun from Zig to Rust manually would take 3 experienced engineers about a year. Using Claude agents, they migrated 𝟓𝟑𝟓,𝟎𝟎𝟎+ 𝐥𝐢𝐧𝐞𝐬 𝐨𝐟 𝐙𝐢𝐠 𝐢𝐧 𝟏𝟏 𝐝𝐚𝐲𝐬, with roughly $165,000 in model usage. Not every project will be like this. Some have dependencies, integrations, and a lot of coordination that still take real time. But for work that’s well-defined, verifiable, and can be parallelized, agents can completely change the math. Maybe that’s the new variable in software estimation: how much of the work can be broken into independent, 𝐯𝐞𝐫𝐢𝐟𝐢𝐚𝐛𝐥𝐞 tasks and run in parallel by agents? A project that looks like years of engineering effort can look very different once you answer that.