Handoff tax
Handoff tax is the price you pay every time one agent passes work to another agent. Most of the cost comes from losing shared context in the transfer.
What the tax is made of
Section titled “What the tax is made of”| Cost | Why it happens |
|---|---|
| Lost context, so worse answers | The orchestrator holds the full state (cart, preferences, history). A subagent only sees the brief it is given, so it works from a summary. |
| Tokens | The subagent needs its own system prompt, tools, and brief, and its answer is read again by the orchestrator. Each handoff can cost several times the tokens. |
| Latency | Extra model calls run one after another while the user waits. Each handoff can add seconds. |
There is also a hidden cost: domains rarely separate cleanly. A returns flow needs order history, the cart, and the catalog. A subagent-per-domain design either copies that access into every subagent, or hands off mid-task and pays the tax again.
How skills avoid it
Section titled “How skills avoid it”A skill is a set of instructions loaded into the agent that already holds the conversation. Nothing is summarized or sent elsewhere, so nothing is lost. You still get per-domain modularity, because each domain’s instructions load only when needed.
When the tax is worth paying
Section titled “When the tax is worth paying”When the subagent’s work is narrow and self-contained, and only a short answer needs to come back. Example: a deep-research subagent that searches, reads, and hits dead ends in its own context, then returns a clean answer.
Delegation is not the same as a hand-off
Section titled “Delegation is not the same as a hand-off”- Delegation: the orchestrator sends a subagent out and back within one turn. This is where the tax applies, and you pay it on every exchange.
- Ownership transfer: a separate agent (for example, one with its own compliance rules) takes over the conversation with the user. You pay once, not every turn.
Takeaway
Section titled “Takeaway”For a tightly coupled, multi-turn conversation that depends on shared state, use one agent with skills. Split work out to a subagent only when the task is self-contained. As models handle more context, the reasons to split get weaker while the tax stays the same.
Sources
Section titled “Sources”- A guide to the anatomy of effective commerce agents, Anthropic (Ali Shazal, Matthew Koen), Sept 2026. Section “Skills, not subagents”.
- Building effective agents, Anthropic, Dec 2024. Routing and orchestrator-workers patterns.