AGENTIC BEHAVIOR

[IMPERA88]

Hello, I have just started experimenting with Haystack. Is there anyway to influence the Agent behavior?
Specifically,

1. is it possible to ask an agent to skip the utilization of a tool even when this is assigned to the Agent?
2. Is the prompt the only way to influence the agentic behavior?
3. Is there any way to 'see' the reasoning of the agent?

Thanks for your support.

[Amnah199]

Hi @IMPERA88, welcome to the Haystack community! 👋
Regarding your questions,

1. For now, you can best guide the Agent's behavior through prompt engineering—specifically by instructing it when to avoid using tools based on your use case.

2. Besides prompt, you can fine-tune the Agent’s behavior using the initialization parameters:

• max_agent_steps – Limits the number of reasoning iterations the Agent performs.
• exit_conditions – Specifies the conditions under which the Agent should stop.
• raise_on_tool_invocation_failure – Controls whether the pipeline should raise an error or continue when a tool call fails.

3. To observe the Agent’s reasoning process, Haystack supports streaming tool_calls and results. The quickest way to inspect this is by using print_streaming_chunk. You can find more details in the [Haystack 2.14.0 release notes](https://haystack.deepset.ai/release-notes/2.14.0).

Let us know how it works out!