Summary
Katelyn Lesse (head of platform engineering at Anthropic) and Angela Jiang (head of platform product at Anthropic) present original research at the AI Engineer conference, introducing the concept that tokens in agentic systems should have specialized roles rather than all performing the same undifferentiated execution work.
The talk defines four token strategies: Execute (standard task completion), Advise (a second role that the executor can consult mid-task), Grade (an evaluator that scores against a rubric and triggers reruns until a quality threshold is met), and Dream (a reflection role that reads executor transcripts and writes learnings to memory for the next run). Each strategy is illustrated with concrete use cases — a sales agent with an adviser checking deal health, a customer service agent with a grader enforcing refund-policy accuracy, a recruiting agent with a dreamer that improves candidate-fit over time.
The research results are specific: on a financial analysis benchmark designed to replicate expert human analyst work, an Execute strategy at a fixed 600,000-token budget reached 76% accuracy, while an Advise strategy at the same budget reached 89%. Crucially, when reframed for real-world consequences — where anything less than 100% accuracy on a P&L is a failed run — the gaps between strategies widen significantly. Lesse and Jiang argue this demonstrates that how you allocate tokens across roles, not just total budget size, produces measurable accuracy gains for high-stakes agentic tasks.
📺 Source: AI Engineer · Published September 14, 2026
🏷️ Format: Keynote Launch
