Inconsistent Agent Responses and Tool Selection in ADK — How to Enforce Retry or Recovery?

[hitflame]

Title: Inconsistent Agent Responses and Tool Selection in ADK — How to Enforce Retry or Recovery?

Body:

Hi team,

I’ve been working with ADK agents and noticed some inconsistent behavior during execution. Specifically:

• In some cases, the agent’s response does not follow the given instructions.
• In other cases, the agent fails to select or invoke the appropriate tool when it should.

In these situations, I would like to have better control over the agent’s behavior—particularly the ability to trigger a retry or force the agent to “recover” and attempt the task again according to the original instructions.

My questions are:

1. Is there a recommended way to enforce stricter adherence to instructions in ADK agents?
2. Can we programmatically detect when the agent fails to follow instructions or skips tool usage?
3. What mechanisms are available to trigger a retry or re-run of the agent in such cases?
4. Are there best practices for guiding the agent to consistently choose tools when needed?

Any guidance or examples would be greatly appreciated.

Thanks!

[surajksharma07]

@hitflame For forcing tool selection, try setting FunctionCallingConfigMode.ANY in generate_content_config — that instructs the model to always pick one of the available tools instead of answering directly.

For retry and recovery, ReflectAndRetryToolPlugin (available via google.adk.plugins) is built exactly for this — it intercepts tool failures, feeds structured guidance back to the model, and retries up to a configurable max_retries.

There's a working sample at contributing/samples/plugin_reflect_tool_retry/ worth checking out.

You can also use after_model_callback on your LlmAgent to programmatically inspect each LLM response and catch cases where an expected tool call is missing before the agent proceeds.

Could you give these a try and let us know if they address the inconsistency you're seeing?