An agent that can only act alone either overreaches or gets stuck. Three tools let you involve someone else, and they differ in one respect that determines which you want: whether the agent retains control of the work.
The agent hands the conversation over and steps out. You configure whether it goes to a human or another AI agent, and a context message accompanies it, explaining where things stand. (If you route human transfers to a specific team or person, confirm that's exposed in your setup — see flag above.)
Write that context message carefully. It's what the person sees first, and a handover that just says "user needs help" wastes the work the agent already did.
The agent gives a task to another agent, waits, and carries on with the result. This is how you compose specialists instead of building one agent that does everything: a front-line agent that delegates order lookups to an orders agent stays simple, and the orders agent can be reused elsewhere.
Two settings. You pick which agent to delegate to — only published agents can be chosen — and a context mode:
Task only — the other agent gets just the instruction. Cleaner and cheaper, and the usual choice.
Full conversation — it gets the whole thread. Necessary when the task can't be understood without what came before, at the cost of a larger, noisier context.
The agent stops and asks a teammate for help: look something up, approve something, do something it can't. It waits, and continues with the answer once it arrives. The tool has no settings — once added, the agent decides when to raise a task and what to ask for, and the run shows a "Waiting for human input" status until a teammate resolves it.
Use it when the missing piece is knowledge or authority rather than the whole task. The agent stays responsible for finishing.
The structured version. Instead of an open question, the teammate gets a typed decision question and picks between named options. It supports multi-step cards with conditional follow-ups — e.g., "Add to roadmap / Implement / Drop," and if they pick Implement, a follow-up asks which approach.
Prefer this whenever the possible answers are known. "Refund in full, refund shipping only, or decline?" is answered in a second and gives the agent something unambiguous to act on; the same question asked openly produces a paragraph it then has to interpret. For open-ended asks with no predefined options, use Request human input instead.
The agent shouldn't be doing this at all — Transfer.
Part of the job belongs to a specialist — Delegate.
It needs something only a person knows — Request human input.
It needs a person to choose between known options — Ask for a decision.
