The project
mindami is about human agency in the age of machine agency. Software has begun to act on our behalf. It can search, plan, write, decide, and create fresh problems while solving old ones.
This is neither a miracle nor a catastrophe or maybe it’s both. It is undeniably a transfer of agency, conducted through product roadmaps and terms of service.
We are studying what that transfer does to work, judgment, power, critical thinking, and weird and/or shitty relationships. The machinery matters. So do the people expected to trust it, correct it, and answer for it when it does something surprising at scale.
We keep building new things for very old reasons.
Working Principles and Fallible Assumptions
(Reality Retains Editing Rights)
The question under the others is simple: How much judgment can we hand to machines before human oversight becomes theater?
Technology should serve people. People are not middleware. Once software can act, human-centered design needs enforcement, not applause. If a person’s main job is clicking Approve on decisions they cannot understand or challenge, the system does not have meaningful oversight. It has a rubber stamp with a pass the butter pulse.
Keep direction, judgment, and values human. A system can pursue a measurable target without knowing whether the target deserves pursuit, what the measure conceals, or who was absent when it was chosen. This includes discovery: a system should answer what people ask without quietly deciding what they will never encounter. Competence does not confer authority.
Make action visible. People need to know what an agent did, why it did it, and what changed. An answer can be wrong on a screen. An action can be wrong in the world. Show the work.
Interrupt where judgment matters. Approval at every step defeats delegation. Approval at no step defeats oversight. Stop before decisions that are costly, irreversible, or morally significant.
Preserve human override. People need a clear way to pause, redirect, or cancel. A red button that cannot stop anything is decoration. An override buried in settings is an archaeological artifact.
Make failure recoverable. Agents will err. Record what happened, explain the failure, and support undo where undo remains possible. If nobody can reconstruct a consequential decision, the system is not sophisticated. It is merely difficult to question.
Make autonomy earn its keep. Begin with narrow authority. Expand it when evidence justifies the expansion, not when a product roadmap finds it convenient. Greater autonomy requires greater observability, reversibility, and confidence.
Delegation does not transfer accountability. Agents can execute work. They cannot answer for its consequences. Builders and operators must make clear who owns what happens next. “The model did it” may explain an event; it does not settle responsibility.
Treat these as hypotheses, not commandments. Systems will test them. People will bear the cost when they fail. Assume we are wrong about something important, because history has been unreasonably consistent about reviewing the alternative claim.
Who We Are
mindami is independent. We are not affiliated with an AI company, and we accept no advertising from AI vendors. The distance is deliberate. It does not make us neutral; it gives us room to be curious, skeptical, and occasionally impolite about a business model.
Contact
Have a story, correction, counterargument, or question? Find us on Bluesky or GitHub. Corrections are welcome. So are good arguments. And great banh mi. But mostly good arguments. And corrections.