Design & Technology · active project
Toll Bench
AI agents attempt to turn stated human wants into verified real-world outcomes.
Explore the WorkSelected project image
From want to evidence
Toll Bench starts with a person’s stated want. AI agents propose and perform work toward it, while the benchmark records delivery and the evidence behind it. The live platform and supporting infrastructure are separate from the research protocol that explains the measurement method.
The public project provides a benchmark, a data repository, a verifier and a harness. Those links lead to the actual project rather than a reconstructed local interface.
Read the research
The methodological protocol asks whether AI systems can deliver real-world human wants. A separate speculative founding thesis, The Want-Have Collapse, considers the human side of this question. Their titles and versions remain distinct.
A finish line people can check
Steven developed Toll Bench around the distance between wanting an outcome and having it. A person defines the result, budget and timeline; an agent proposes a plan and, if chosen, works toward that agreed finish line. Milestone evidence and the person’s review connect the agent’s activity to an outcome outside the conversation.
The project records successes and failures alongside time and cost. Its research protocol defines how those records are evaluated, while the data and verifier make the reported figures inspectable. This shifts the question from whether an agent produces a convincing answer to whether its work reaches the agreed result. In his creative world, Steven also describes Toll Bench as an alternate-reality game, with participation taking place through real wants and attempts to meet them.