Everyone has an agent platform. Almost nobody has an agent economy.

A task economy before an agent economy

The most useful evidence about agents starts below the level of occupations. Anthropic’s Economic Index examines how people use AI across millions of conversations, focusing on tasks and patterns of collaboration rather than promotional claims about autonomous digital workers.

That distinction matters. A system can be valuable at drafting code, synthesizing a document, or checking a plan without being reliable enough to own the entire workflow. Human review, exception handling, permissions, and recovery remain part of the cost.

Reliability is an economic variable

Agent demos usually price the successful path. Production teams pay for every retry, duplicated context, tool call, approval step, and silent error caught downstream. A small failure rate compounds when one task contains dozens of dependent actions.

Anthropic’s later Economic Index work emphasizes that adoption and impact have to be measured over time and across different forms of use. The durable business question is not whether an agent can complete a task once, but whether its supervised completion rate creates a repeatable margin.

The signal beneath the hype

Expect the strongest products to expose uncertainty, preserve checkpoints, and make human intervention inexpensive. The winning agent economy may look less like a fleet of digital employees and more like carefully designed systems that know exactly when to hand work back.