The figures come from OpenRouter, a service that routes requests to AI models, as charted by venture firm Andreessen Horowitz. Since agents overtook humans on the platform in February, agent token use has grown 14x while human use grew 2.8x. OpenRouter classifies each API key as agentic, mixed or human using a weighted score that includes tool call rate, turn count and timing between messages.

The catch is that more than 85% of agent tokens come from cached prompts, meaning the agents are mostly rereading context they have already processed. Cached tokens cost far less than fresh processing, but they still occupy memory, and a16z links the trend to rising demand for high-bandwidth memory. The same pattern shows up elsewhere: a call centre consultancy testing DeepSeek on rented Nvidia H200s found that 96% of its agents' input on Claude Code in September was re-reading old conversation.

Caveats apply. The data covers one platform only, the trend is not perfectly smooth with dips in April and July, and it measures token volume rather than spending. Still, agent adoption is broadening beyond OpenRouter: McKinsey's 2026 State of AI survey found 40% of respondents from large organisations scaling AI agents, up from 27% a year earlier.