Agent Script for deterministic multi-agent routing
Key Takeaway
Agent Script lets you make multi-agent routing deterministic instead of relying on the LLM to follow instructions. Context doesn't flow automatically between agents, so a connected subagent gets a value only if its inputs block maps it.
Multi-Agent Orchestration lets one orchestrator agent route work to other complete agents, called connected subagents. You edit Agent Script in an agent's Script view. With Agent Script, you can force a route when a condition is true, hide options from the LLM, pass specific data between agents, and chain agents after a reply. Prompt instructions alone can't guarantee this control, because the LLM only usually follows instructions.
Orchestrator agents and connected subagents
An orchestrator agent handles the first conversation with the user, then routes and delegates tasks. A connected subagent is an entirely separate agent in your org that the orchestrator uses like a subagent. Unlike a standard subagent, which is only instructions and actions inside one agent definition, a connected subagent is a complete, independent agent with its own subagents.
A multi-agent design fits when one agent is no longer enough. For example, an Instructions block may have grown to cover many domains and routing keeps breaking. Or users may have to choose between several separate agents and repeat themselves when they switch. The orchestrator then acts as a single front door for all of those agents.
There is no system limit on the number of connected subagents. Internal testing found no latency impact with one orchestrator and 8 connected subagents, each with 8 subagents. Even so, multi-agent latency is always higher than single-agent latency.