Is there any way to force the agent to continue processing or retry?

[hung12ct]

Currently, sometimes after the ADK gets the result from the tool, it stops and doesn’t continue thinking. Is there any way to check if the event is from a tool so that the agent should continue running? I’m using ADK 1.10 and Gemini-2.5-pro. Thank you

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: The behavior you're observing is expected in the ADK's event-driven architecture. To control the agent's execution flow and ensure it continues processing after a tool call, you can use callbacks, specialized tools for long-running tasks, and workflow agents like LoopAgent.

---

Hello! It seems you're looking for a way to ensure your agent continues processing after a tool returns a result. The agent pausing is part of the ADK's asynchronous, event-driven design, where it waits for the next event after a tool's execution. [3] Here are a few strategies to manage this and force the agent to continue its thinking process:

1. Use Callbacks for Fine-Grained Control

Callbacks allow you to hook into the agent's execution lifecycle and modify its behavior. You can use a before_model callback to inspect the conversation history and the last tool output, and then modify the prompt to guide the agent on what to do next.

For example, you could check if the last event was a tool output and, if so, add a specific instruction to the prompt to force the agent to analyze the result and decide on the next step. You can find more details on how to implement callbacks in the official documentation. [5]

2. Handle Long-Running Tasks with LongRunningFunctionTool

If your tool is performing a task that takes a significant amount of time, the agent might be pausing while waiting for the result. For such scenarios, it is recommended to use the LongRunningFunctionTool. This tool allows the agent runner to pause the agent's run, and your client code can decide whether to wait for the result or continue with other tasks. [2]

3. Implement a Workflow with LoopAgent

For more complex scenarios where you need to retry a task or process something in a loop until a certain condition is met, you can use a LoopAgent. This agent can contain other sub-agents and execute them in a loop. You can create a condition-checking agent that determines whether the loop should continue or terminate based on the state, which can be updated by your tools. This gives you a structured way to enforce a continuous processing flow. [4]

4. Manage State with InvocationContext

The InvocationContext (ctx) provides access to the agent's execution state. You can use ctx.session.state to store information across different turns of the conversation and use this state within your callbacks or tools to make decisions about the agent's behavior. For instance, a callback could check a flag in the state to decide whether to end the invocation or prompt the agent to continue. [6]

By combining these strategies, you can gain more control over your agent's execution and ensure it behaves as you expect after a tool call.

---

[1] https://google.github.io/adk-docs/get-started/about/
[2] https://google.github.io/adk-docs/tools/function-tools/
[3] https://google.github.io/adk-docs/runtime/
[4] https://google.github.io/adk-docs/agents/multi-agents/
[5] https://google.github.io/adk-docs/callbacks/
[6] https://google.github.io/adk-docs/context/