Pre-routing tools and agents before ADK execution

[sauravsingla]

I’m exploring a pattern where the model or agent runtime does not receive every available capability by default.

Instead:

available tools / agents
→ policy + scope filtering
→ optional task-aware routing
→ smaller selected set
→ normal ADK execution

The goal is not to replace orchestration, but to make capability selection explicit, reproducible, and observable before execution.

I’ve implemented this experimentally in AgentWeave and would be interested in feedback from ADK users:

Where do you currently perform tool/agent candidate reduction in larger systems?

Repo: https://github.com/sauravsingla/agentweave

[surajksharma07]

Worth knowing ADK already covers this natively: tool_filter/ToolPredicate on toolsets for tool-level filtering, before_tool_callback for argument-level policy and sub_agents/transfer_to_agent for agent-level routing. Closing this as Show and Tell since it centers on your own external project rather than an open ADK question, keep the ideas coming.