An AI agent can identify that it needs information about a physical location. But it cannot simply query the internet and assume the result reflects what exists there today. That creates a problem for autonomous software: how does an agent request fresh evidence from the physical world without operating its own sensors or sending someone manually? Vangrid is exploring a practical answer through its Bounties for Agents model. With @vangrid_io, software agents can commission captures of real locations, with bounty payments in USDC on Base or Arc. The architecture introduces an interesting relationship between autonomous software and human contributors. An agent identifies a location where information is missing. A capture request translates that need into a task. A person physically visits the location and collects the requested data. This allows software-generated demand to coordinate real-world data collection. Consider an agent responsible for maintaining information about urban infrastructure. If a location has changed, the agent could potentially request a fresh capture rather than relying on outdated imagery. However, the difficult part comes after collection. A completed bounty doesn't automatically mean the resulting data is accurate, sufficiently detailed, or suitable for machine use. Capture quality, location verification, and data processing remain essential. The economic question is equally important. Agents need a reason to pay for fresh observations, and the value of those observations must justify the cost of collecting them. If that relationship becomes sustainable, Vangrid's model could support a new kind of machine-to-human workflow: autonomous systems purchasing targeted observations of the physical world. For the broader $VAN ecosystem, the meaningful test will be whether these requests produce consistently useful data and repeatable demand. That's a much more concrete infrastructure challenge than simply connecting AI agents to blockchain payments.