Supervisor topologies pay one extra LLM call per hop because the supervisor must read every worker output and decide what runs next; swarms collapse this by letting the worker emit the next handoff tool call directly.
Picture a manager who insists on personally approving every step of every project. A worker finishes a task and reports to the manager; the manager reads the report and tells them what to do next; the worker does it and reports again; repeat. The manager's reading and deciding is real work, but it doubles the back and forth on every step. Now picture a team where workers just decide among themselves who handles the next step. The decisions still get made, but no manager turn is inserted between worker turns. The first setup is the supervisor topology and its tax is the manager's per-hop call. The second is the swarm, and it saves the manager call by letting the workers route directly.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
The supervisor topology is one of the most common multi-agent patterns in 2026 production systems, and also one of the most commonly under-costed. Engineers see the centralised-routing pattern, understand it gives them an obvious place to put policy and observability, and ship without quantifying the per-hop tax. The cost shows up later as monthly invoice growth that scales unexpectedly with workflow hop count.
The tax is structural. The supervisor turn is a full LLM call sandwiched between every pair of worker calls. For N hops, the supervisor topology costs 2N LLM calls where the swarm costs N. The 2x multiplier is not a tuning artefact; it is the architectural consequence of separating routing from execution.
This deep dive walks the mechanics that produce the tax, dismantles the three distractor options as configurable knobs rather than structural properties, names the workflows where the auditability premium justifies the cost, explains the cheap-supervisor pattern that contains dollar cost but not latency, and lays out the hybrid pattern that combines swarm and supervisor for the best of both.
The mechanics: why the tax is exactly 2x per hop
Walk one hop in a supervisor topology end to end. A worker is invoked with the current state. The worker calls its LLM, does substantive work (calls tools, produces output), and finishes its turn. The framework writes the worker's output to shared state. The supervisor agent fires next: its prompt is the routing prompt, its input includes the conversation context and the new worker output, and its LLM call produces the next-speaker decision. The framework reads the supervisor's decision and invokes the next worker. The next hop begins.
The critical detail is that the supervisor call is serial with respect to the worker calls. It must complete before the next worker can start, because its output IS the next-worker selection. Per hop, you pay two LLM calls (worker + supervisor) in series.
Now walk one hop in a swarm topology. A worker is invoked with the current state. The worker calls its LLM, does substantive work, and emits a handoff tool call (transfer_to_research_agent(brief)) as the last action of its turn. The framework intercepts the tool call, recognises the transfer prefix, swaps the active agent, seeds the new agent with the brief, and the next hop begins. The substantive work and the routing decision are produced by the same LLM call. Per hop, you pay one LLM call.
The delta is exactly 1x extra LLM call per hop, and it scales linearly. A 10-hop workflow is 20 calls supervisor versus 10 calls swarm. A 30-hop workflow is 60 versus 30. Latency tracks the same way because the supervisor call is serial: wall-clock time roughly doubles in the supervisor topology compared to swarm for the same hop count.
This is what the question means by "structural" cost. The 2x multiplier comes from the topology's separation of routing into its own turn, not from any specific model choice or context configuration. Any supervisor topology pays this tax; no amount of model tuning eliminates it.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- OpenAI's Agents SDK handoff as tool call pattern is the canonical 2026 swarm implementation that avoids the supervisor tax.
- LangGraph's create_supervisor and create_swarm prebuilts let you pick either topology with one function call, making the cost versus control trade-off explicit.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhen does the cost premium of supervisor topology earn its slot despite the 2x per-hop overhead?
Workflows that need centralised routing decisions for auditability, compliance, security, or complex multi-criteria routing logic. The supervisor's call becomes a policy-enforcement point that worker-distributed routing cannot match. Examples: regulated industries that must log every routing decision, security contexts where worker agents cannot be trusted to route safely, multi-criteria routing that needs to consider state beyond what any single worker sees.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Identifying the supervisor's cost premium as a model-tier or context-window choice (configurable knobs) rather than as the structural per-hop routing call that the topology forces by design.
60 second bullets to scan on the way to the call.
Compare LLM calls per hop in supervisor versus swarm topology
Explain how latency tracks LLM-call count in both topologies
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.