“A recommended best practice for building reliable AI agents is to limit an LLM's access to approximately five tools at any given step in a workflow.”