Loading tools dynamically as JSON at runtime ?

[glc-froussel]

Hello everyone,

From the documentation (https://strandsagents.com/latest/documentation/docs/user-guide/concepts/tools/#building-loading-tools) I understand that we have 2 ways to load tools for an agent : as function or as python module

Is there any way to load tools "dynamically" at runtime directly with a JSON instead ? I have all of my tools configs currently stored and versioned outside the application, and that allow me to publish new tool configs without deploying code (and even A/B testing different version of the same tools dynamically).

I was something I can do if I call directly the LLM provider (like converse API from Bedrock or OpenAI API) but I couldn't find a way to do it with Strand Agents.

Any idea ?

[zastrowm]

Two related notes:

• Agent has a load_tools_from_directory to dynamically load tools at runtime, though it's disabled by default. Unfortunately I can't find our user-guide on it
• AgentConfig is an experimental feature for defining agents via JSON and something we're soliciting input on (CC @Unshure)

[glc-froussel]

Thanks for the quick feedback @zastrowm !

I found this section in the user guide:
https://strandsagents.com/latest/documentation/docs/user-guide/concepts/tools/#auto-loading-and-reloading-tools
From my understanding, it only supports module-based tools, which means I would need to generate a Python file from my JSON configuration.

And regarding AgentConfig, it looks very promising! According to the documentation, the tools attribute also expects module-based or function-based tools. It could be a great enhancement to support direct JSON-based configuration as well.