Documentation
¶
Overview ¶
Package router assembles the top-level agent for a workload: a Chat-mode coordinator with the workload's specialists as SubAgents.
This uses ADK v2's canonical dispatch pattern (see google.golang.org/adk/v2/examples/multiagent/task_sub_agent):
- Chat-mode LlmAgent = the coordinator (top-level driver).
- Task-mode LlmAgents in SubAgents = per-failure-mode specialists. ADK auto-installs one `task` tool per Task sub-agent; the coordinator's LLM invokes a specialist by calling that tool.
- SingleTurn-mode LlmAgents in SubAgents = cheap classifiers. ADK auto-installs one `single_turn` tool per SingleTurn sub-agent; the coordinator's LLM invokes the classifier by calling that tool.
This is "LLM-as-router" delivered via the SubAgents dispatch pattern rather than an explicit workflow.Workflow graph. runner.Runner requires the root agent to be a Chat-mode LlmAgent (see runner.go isLlmAgent + Mode check), so a bare Workflow cannot be a root agent — a graph shape must be composed underneath a coordinator or driven out-of-runner.
Explicit workflow.Workflow shapes (fan-out-fan-in, adversarial verifier, autonomous loop, etc.) remain available for use cases the SubAgents pattern doesn't cover; those land as a follow-on when the first shape needs them.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
Types ¶
type Config ¶
type Config struct {
// Bundle is the workload definition. Used for the coordinator's
// name, description, and roster ordering.
Bundle workload.Bundle
// Specialists is the roster of specialist agents, indexed by
// spec name. Every name in Bundle.Specialists must be present.
Specialists map[string]adkagent.Agent
// Model is the model the coordinator itself uses. Specialists may
// use different models (their construction already bound them).
Model model.LLM
// Instruction, when non-empty, replaces the default coordinator
// system prompt.
Instruction string
}
Config describes how to build a coordinator for a given workload.
Source Files