Can trustworthy autonomy emerge from the harness instead of a prescribed reasoning schema?
Open notebook5 open questions
Ideas in progress
Questions I have not resolved yet. Publishing them early makes the edges of the work visible.
What is the right abstraction boundary between a skill and a durable autonomous workflow?
Can rollout intelligence transfer across applications without erasing local risk policy?
What remains defensible product value when frontier models can construct their own orchestration?
How do we evaluate an agent's decision when the production counterfactual is unknowable?