Discussion on Agent component

[bilgeyucel]

This is the discussion board for the experimental Agent component

Agent is a new experimental component that implements tool-calling functionality with provider-agnostic chat model support. It can be used either as a standalone component or within a pipeline.

💻 See the implementation details here
🧑‍🍳 See the Agent in action in Cookbook: Build a GitHub Issue Resolver Agent

Note: Experimental features live in this repository for a fixed period of time. We don't guarantee that we will continue maintaining experimental features. But if they are successful, stable and if you like it, we will move the feature to the core Haystack package.

[d-kleine]

I just took a look into the cookbook example - unfortunately, this exceeds my private tier 1 rate limit for Anthropic (This request would exceed the rate limit for your organization). Nevertheless, I would like to share some observations and thoughts on the implementation:

Implementation:

• The Agent() implementation looks very intuitive to me, particularly its clean abstraction for tool integration and being provider-agnostic
• Probably more related to the Tools: About the Agent()'s tools, I wondered also whether it's also possible to call tools simultaneously (not sequentially) or with a given prioritization within the agent
• Maybe it would be helpful to have a direct monitoring feature to log and optimize the agent's behavior over time?

Cookbook:

• I really enjoyed exploring the cookbook use case for custom agent implementation! But as a user, I found it to be excessively long. For beginners using the Agent() component, a simpler "quickstart" focusing just on basic agent/tool interaction could be more helpful before tackling the more advanced custom features
• Not directly related to the Agent() implementation itself, but in the cookbook example, the nested result access tool_call_result.result confused me

[ernoaapa]

Any plans on adding support for AsyncPipeline? Would be great to get the progress of Agent executing tools before getting the final result.

[sjrl]

We definitely have plans to look at supporting streaming from Tool Calls (see issue here) which I believe should cover your request. As for using AsyncPipeline I'll bring it up with the team.

[d-kleine]

This idea might go beyond the current scope of Haystack, but I would be curious to know if there are any plans to incorporate learning capabilities (for example via reinforcement learning, e.g. GRPO) into agentic workflows?

The current agent implementation seems to be static - once configured, its decision-making process doesn't evolve based on outcomes. What I have thought of is making agents dynamically learn and improve through experience, adapting their tool selection and reasoning strategies over time. From my POV, this could make agents smarter about which tools to use when, help them continuously improve (without manual retraining), and make better use of resources. They could also optimize for whatever metrics matter most in your specific domain, not just a generic performance. For example, in a customer support scenario, agents could learn which retrieval tools and metrics work best for technical vs. billing questions.

Also, in a multi-agent pipeline, each agent might be able to learn its specialized role in the workflow, developing expertise in specific tasks while understanding how to effectively hand off information in the workflow.

Have you already considered this direction? What do you think about this out-of-the-box idea? 🙂

[julian-risch]

These are great ideas for extending the Agent implementation. We're excited about the idea of an adaptive tool selection. I very first step in that direction is deepset-ai/haystack#9043 Giving the Agent access to a large set of tools or better to say a way to look up tools when needed.

[julian-risch]

With the Haystack 2.12 release, we completed the Agent experiment and here is the new docs page. We'll add a tutorial very soon!

[bilgeyucel]

Tutorial is out, explaining how to create a web search agent Build a Tool-Calling Agent