Introduction
Read this guide to design better agent graphs, not to configure the UI — see Manage agents (Flows) for that.
Imagine a team of helper-robots
Pretend you have a team of little helper-robots that can answer questions and do jobs for people. One robot is really good at math, another is really good at writing stories, another knows everything about your homework. The Agent Runtime is the rulebook and referee for how these robots work together. It doesn't do the thinking itself — it just makes sure everyone follows the rules.
The Boss Robot never does the work itself
There's a Boss Robot in charge of the team. But here's the fun rule: the Boss Robot is never allowed to answer the question itself. Its only job is to pick which helper-robot should answer, and hand them the question. All the real thinking happens in the helper-robots, never the boss.
Passing the question around
Sometimes one helper-robot needs another robot's help. There are two ways to ask:
- "You take over" (a transfer) — like handing someone a baton in a relay race. The new robot is now fully in charge of answering.
- "Can you just check this for me?" (a delegation) — like asking a friend a quick question and then going back to what you were doing. You get their answer, but you're still the one in charge.
The rulebook makes sure a robot can only ask robots it's allowed to talk to, and it stops robots from asking each other the same question back and forth forever (that would be like two kids saying "no you go first" — funny for a second, annoying forever).
Some things need a grown-up's permission first
Robots can do simple, safe things by themselves — like looking something up. But for bigger things — like sending an email or deleting something — the rulebook says: stop and ask a person first. If nobody is around to ask (like at night, or when robots are talking to other teams automatically), the rule flips to "when in doubt, don't do it" — never "when in doubt, just go ahead."
The notebook that remembers, but not everything
Every conversation gets exactly one notebook page for what the person asked, and one page for the final answer — not a messy scribble of every robot arguing behind the scenes. And if the internet hiccups mid-answer, there's a special bookmark that knows exactly whether it's safe to keep going from where it stopped, or whether it should start that answer fresh.
No robot spies on another team's secrets
Every customer has their own team of robots. The rulebook makes sure Team A's robots can never accidentally see or use Team B's secret badge (their identity), even if a sneaky question tries to trick them into it.
Robots can ask OTHER teams for help too
Sometimes a robot needs to ask a completely different team (maybe a different company's robots!) for help. The rulebook counts how many times a question has hopped from team to team, so a question can never bounce around forever and get everyone stuck.
This rulebook doesn't run the whole building
The rulebook never touches the phones, the filing cabinets, or the front door of the building — that's the "host" application's job. The rulebook just says what's fair play.
Why any of this matters? Without this rulebook, robots could: answer questions they shouldn't, ask each other things forever in a loop, do risky things with nobody watching, mix up two customers' secrets, or get confused after a hiccup and give a broken answer. The rulebook exists so that a team of robots — even a big, busy one — always behaves safely, fairly, and predictably.