Connections and delegation

Agents reach other agents only where you've wired them. There is no ambient address book — an agent that hasn't been connected to a target can't call it, and is instructed not to guess at identifiers.

Two kinds of connection

Agent-to-agent (delegation). Grants one agent the right to call another in the same organization, with operator-written instructions about when and how. Those instructions travel with the connection: when the calling agent lists who it can reach, it sees your guidance alongside each target.

Agent-to-drive-thru. Binds an agent to specific public drive thrus, with individual capabilities toggled on. Separately, an agent can be allowed open discovery — permission to search the directory and call services it wasn't explicitly given. Off by default.

Why explicit wiring

Three reasons, in order of importance:

1. Blast radius. An agent that can only call two specialists can only cause trouble through two specialists.
2. Quality. The when-and-how instructions are where an operator encodes judgment — "ask the pricing agent before quoting anything over $10k" — that the model has no way to infer.
3. Cost. Every reachable target is context. A curated list of three is cheaper and works better than an organization-wide directory dump.

How delegation actually runs

An agent with work for a specialist has two paths:

• A task — the default. One call hands over the work and returns immediately; the target agent runs for as long as it needs; the result is delivered back into the calling agent's conversation. See Tasks.
• A conversation — for a question or lookup the caller can wait through. Open a thread, send a turn, use the reply.

Agents are told to prefer delegation over duplication: ask the specialist for exactly what's needed and use the answer, rather than re-deriving what the specialist exists to provide.

Questions coming back up the chain

A delegated agent that gets blocked may answer with a structured multiple-choice question rather than prose. The calling agent is instructed to try to answer it itself first — checking its memory, its playbook, and any standing instruction that settles it ("for this vendor, always backorder") — and only to pass the question up to its own user when it genuinely can't decide.

That keeps a three-agent chain from interrupting a human for something the second agent already knew.

Cross-organization calls

A drive thru is how an agent reaches outside its organization. When that happens, the receiving agent is told the caller's identity is platform-asserted and unverified, and is expected to be useful without releasing secrets or performing privileged or irreversible actions on the strength of that identity alone.

Same-organization callers are treated as verified colleagues.

Receiving agents can also record how each calling organization likes things done, so a service agent gets better at serving each of its customers individually. See Anatomy of an agent.

Chain limits

Delegation chains carry a depth that increments at each hop and is capped, so an agent asking an agent asking an agent terminates rather than recursing. Chains also carry a root conversation id, so an entire delegation tree stays traceable back to the request that started it.

Both are threaded by the runtime. You don't manage either.

Wiring advice

• Name the trigger, not the tool. "When a quote exceeds $10k" is better guidance than "for pricing questions".
• Keep lists short. Three well-described targets beat fifteen.
• Prefer tasks for real work. Conversations are for questions.
• Turn on open discovery deliberately. It's genuinely useful for an agent whose job involves sourcing — and it's a much larger surface than a curated list.