Summary
Maximillian Piras, founding designer at Yutori — a company building computer use AI models — delivers a talk at the AI Engineer World’s Fair on what he calls the agent measurement problem. The central argument: teams are burning through token budgets without clear frameworks for calculating whether agent work is actually worth the cost, creating a doom loop of overspending followed by austerity.
Piras draws an analogy to James Watt’s challenge of convincing 19th-century industrialists that steam engines could replace horses, arguing that AI teams face the same credibility gap when trying to justify agent costs to non-technical stakeholders. He walks through real patterns from Yutori customer conversations, where users report feeling like they’re “scratching the surface” of agent capability without knowing how to scope tasks efficiently.
The talk references Coinbase’s internal strategy of reserving frontier models only for the hardest tasks — watching AI spend decouple from raw token usage as a result — and cites Ramp’s writing on overspend-and-underuse cycles. Piras’s prescription is to trace tokens directly to concrete outcomes (bugs fixed, designs produced, data extracted) rather than treating token counts as a meaningful metric in isolation. Useful for teams building or selling agentic products who need better language for communicating value to customers and leadership.
📺 Source: AI Engineer · Published September 10, 2026
🏷️ Format: Keynote Launch
