Questions
Questions I'm working on
How might groups of agents become resilient or improve over time?
Incentive design and mechanics can allow agents to maintain and act in line with stable intent: we need memory, adversarial pressure, audits and evaluations, learning loops, and permission for them to say "no" or "I don't know."
How should we treat memory or context as a managed resource?
Failure modes where capable reasoning agents break down in prod are often due to not having access to the right priors or relying on the wrong ones rather than raw intelligence.
Where should human authority sit when agents act on our behalf?
The useful taxonomy is not human-in-the-loop versus out-of-the-loop; it is which decisions or actions require taste, consent, accountability, reversibility, or felt empathy before the system is allowed to continue.
How do you assure work you can't fully inspect?
I spent years designing audits that assured thousands of judgments without reading every line. Anyone building with agents has unknowingly been thrust into those same waters — sampling, materiality, burden of proof — but very few of us were ever taught this discipline. How can we help the 99%?
What is a generative agent's contribution actually worth?
Most token economics stops at cost — metering consumption, optimizing spend. The harder problem is the other side of the ledger: valuing what agents contribute, attributing it, and letting that value accrue to the people and systems that produced it.
How can learning systems extend the humans they serve while preserving what makes them distinct?
The work is to build systems that notice patterns, reduce load, expand agency, and reflect a person's own strain of creativity and taste without collapsing into surveillance, flattery, dependency, or losing their right to remain surprising.