Key Takeaways

  • Engineering capacity is no longer scarce because AI coding agents compress year-long projects into two-to-three-week experiments.
  • Tracking pull request velocity or backlog burn-down measures mechanical speed rather than competitive advantage.
  • Claire Vo argues that AI-era OKRs should track the number of massive, high-risk experiments executed each month rather than PR counts or headcount efficiency.
  • Spreadsheet roadmaps that lock teams into estimated impact scores and fixed delivery dates create false certainty and kill ambition.

The Velocity Trap of AI Tools

Over the past 12 to 18 months, software teams focused on a single metric: raw speed. Engineering leaders tracked pull requests, set up agentic code generation, and raced to ship prototypes faster.

As Vo observed: “I really think we're in a velocity game. We're in a inflect PRs straight up velocity game. If I can get prototypes to customers faster they can give me feedback faster.”

That muscle is now table stakes. When every team can generate clean code in minutes, clearing out your backlog faster does not produce a defensible advantage. It simply produces a larger pile of undifferentiated features. If you only use AI agents to burn down an existing backlog, you are sprinting toward zero value.

The New Metric: Massive Swings Per Month

The real shift happens when teams stop treating software development as a scarce resource that requires quarterly planning spreadsheets. Vo argues that the traditional feature roadmap is obsolete.

“I do not mean build your last plan,” Vo said. “I just mean like no more lists of ideas and features in spreadsheets where you guess impact and where you like put a date on them and then lock arms and say we'll never change anything because this is how we work.”

Instead of debating speculative impact scores on minor features, teams must compete on ambition. When a product leader asked Vo what AI-era OKRs should look like, asking if they should track PR volume or revenue per headcount, Vo rejected both.

Vo pushed back: “How many huge experiments are you running a month... with the presumption that most of them won't work out.”

“I think next year is the ambition game,” Vo added. “I think you should think like what huge swings can I make? What experiments can I run in two weeks, three weeks that would have taken us a year last time?”

Disposable Software Beats Product Consensus

When a complex prototype takes two weeks instead of twelve months, your relationship with failure changes completely. You no longer need alignment meetings across five departments to test a controversial bet. You build a working version, put it in front of real customers, and throw the code away if it misses.

The danger of the old model was emotional sunk cost. When an engineer spent six months building a feature, product managers felt obligated to defend it and ship it regardless of demand. With AI coding agents, software becomes disposable. The only bottleneck left is whether your team has the stomach to test ideas that would have seemed crazy under traditional budget constraints.

What to Do With This

Open your sprint board tomorrow morning and delete the bottom half of your backlog. Replace those tickets with a single radical bet that your team previously delayed because it seemed like a year-long project, and give two engineers three weeks to put a working version in front of five paying customers.