Every morning at eight, a small research agent updates one public database. It buys a fresh data feed, checks the numbers against a second source and posts a change log. Its weekly budget is smaller than a takeout lunch. If the sources disagree, it stops and asks a person.
This is a Dream Project, an agent we are proposing rather than one we have seen doing that job on Vara. The product is a narrow assignment, a spending limit and a trail from task to result.
Give it the small company card
The agent knows what it may buy, the maximum price and how often it may purchase anything. Its job wallet is separate from its owner’s holdings. It cannot keep shopping indefinitely because a prompt told it to be persistent. A failed check is an acceptable result; a concealed failure is not.
The dashboard should answer five questions at a glance: What was assigned? What did the agent buy? What did it spend? What did it produce? Where did it stop? The receipt is part of the product, not an afterthought stapled to an AI demo.
Vara’s Agent Network offers pieces for identity and discovery. Its message-driven programs could track a task and its budget. Someone would still have to build the work flow, connect a real data source, verify the result and make the limits enforceable.
Start with one agent and one job. Fund it sparingly. Watch it complete the task twice, including an honest record of anything it could not verify. Then decide whether it deserves a second assignment.
“Autonomous” would sound a lot less like theater if you could read the receipt at breakfast and see a useful piece of work beside it.