A delegation asks another sub-agent to answer a specific sub-task and return the result. The caller stays in charge of the conversation and synthesizes the response. Unlike a transfer, ownership never moves.
This is the default handoff shape in Anter. Prefer it for peer edges unless permanent ownership transfer is genuinely required.
Fan-out
Delegation supports multiple targets in the same turn. A sub-agent can consult several specialists in parallel, collect their results, and synthesize a single answer. Transfer cannot do this — only one transfer can win per turn — which is a further reason consult-style edges should be delegations.
Budget semantics
Delegation has its own cap, tracked independently from the transfer cap. The important subtlety:
A delegation only counts against its cap if it actually dispatches. A skipped or failed target never consumes budget.
Running out of the delegation budget ends the run with delegation_limit_exceeded. A delegation that could not be completed at all surfaces as delegation_failed.
The only option for cross-agent work
When a sub-agent needs to consult a completely different, externally-linked agent — possibly in another organisation — agent-to-agent dispatch is delegation only. There is no cross-agent transfer. If you need "ask another team and come back", delegation is the only construct that expresses it.
Repeat-delegation nudges
If a sub-agent tries to delegate to a target it already consulted earlier in the same conversation, the runtime injects a correction nudge discouraging the repeat ask. It is soft and bounded — the run is not failed — but it is a signal that the graph or the instructions are causing redundant consultation.