Experiments, prototypes, and research from the Phoebe team.
1,000 subagents, one shared state, and one big beautiful graph.
Lessons from building (and rebuilding) Phoebe's evals: let customers define success, distrust simulations, and treat user error as a test case.
