I have built an agent system using this architecture: Hierarchical Agent Teams
A hierarchical team is tasked with completing actions. With the introduction of o4-mini the top level supervisor will get into an endless loop, this does not happen using o3-mini of gpt-4.1
for example, a supervisor creates a plan for a worker to fetch some data. the worker fetches that data, adds the message to our internal state, and reports back to the supervisor. the supervisor then makes the same request again as if it did not receive the intended data. the worker then fetches it again—this time the supervisor understands it has received and moves on to the next task. this loop happens again with another tool—sometimes there are many loops and it never resolves.