Economic Primitives — How AI Systems Should Reason About Cost and Value
Anthropic's research team has published a new position paper titled Economic Primitives for AI Systems, providing a structured framework for how AI agents should think about resource expenditure and value creation during extended task execution. As AI systems take on longer-horizon tasks with real-world cost implications — compute spend, API calls, human time — the absence of a principled economic model has led to wasteful and sometimes counterproductive agent behaviour in production.
The four core primitives proposed
- Task value estimation — before beginning a multi-step task, an agent should form a prior estimate of the expected value of completing it, factoring in uncertainty
- Marginal cost accounting — each action should be evaluated against its marginal cost (tokens, wall-clock time, external API credits) and the incremental value it is expected to add
- Abandonment thresholds — agents should have explicit stopping criteria; continuing a task when expected value drops below accumulated cost is economically irrational
- Delegation economics — when spawning sub-agents, the orchestrator should explicitly budget their resource envelope and enforce it
The paper includes worked examples from real Claude Code agentic runs that demonstrate cost overruns, redundant tool calls, and unnecessary sub-agent spawning that the framework would have prevented. Anthropic notes it is evaluating which primitives to encode directly into future model training versus which to implement at the agentic scaffolding layer.