Explicit Prompt for AgentTool Invocation - Enabling Guided Encapsulation

[LukasB97]

Hello ADK Team and Community,

Context & Alignment with Tool Paradigm:

The ADK allows an LlmAgent (Agent A) to use another agent (Agent B) as a tool via AgentTool. This aligns with the tool paradigm where an LLM can invoke external capabilities. Typically, LLMs provide arguments to control a tool's specific action.

The Problem: AgentTool Lacks Dynamic Instruction, Especially for Encapsulated Agents

Currently, when Agent A invokes Agent B via AgentTool, Agent B acts based on its static instruction and the session context provided by the framework. Agent A cannot directly pass dynamic, task-specific instructions to Agent B for that invocation.

This limitation is particularly acute when Agent B is configured for strong encapsulation using include_contents='none'. In this mode:

• Agent B receives no prior conversation history.
• Its execution relies solely on its static instruction and any input passed in the current turn.
• Without a way for Agent A to inject specific instructions, Agent B operates almost completely "blind" to the nuanced requirements of the current task assigned by Agent A.

The workaround of passing task details via shared state is indirect, couples agents implicitly, requires Agent B to be aware of specific state keys which undermines clean separation.

Proposal: Explicit Prompt Injection - Enabling Guided Encapsulation

We propose allowing Agent A's LLM to optionally generate an explicit prompt as part of the AgentTool invocation. The ADK framework would pass this dynamic prompt to Agent B for that specific execution.

Benefits:

• Allows Agent A to provide necessary, task-specific guidance directly to Agent B, even when Agent B is encapsulated (include_contents='none').
• Enables effective use of highly specialized, stateless agents by providing their runtime context dynamically.
• Maintains strong agent encapsulation (Agent B doesn't need global history) while allowing flexible control.
• Aligns AgentTool functionality more closely with the expected control of regular tools.
• Avoids fragile reliance on shared state for passing core task directives.

We believe this is crucial for building robust and modular multi-agent systems with well-encapsulated components.
What are your thoughts?

[GuillermoGaete]

+1

I've posted some similar there: #61 (comment)

[fedorzh]

To add here a specific case. I have an agent which is supposed to have input_schema and output_schema set. It should only use these schemas and the instruction to produce a correct output. However, it does not do so if include_contents is none - it clearly doesn't know about input contents, and does not produce the correct output.