ExperimentProposed topic — outline, not published findings
Reinforcement Learning for Agent Behavior
Exploring whether reward-shaped training can make agents more reliable at multi-step tasks.
The question
Prompting and scaffolding get agents surprisingly far. But can training with task-level rewards make multi-step behaviour more consistent than prompting alone?
What we'd measure
Task completion rate, steps to completion, tool-call error rate, and cost per successful run — against a prompted baseline on the same evaluation set.
Status
An area of active exploration. No results to report yet; we'll publish methodology and findings here when we have them.
