[FEATURE] Strands Agents should provide a built-in intent confidence

[leecooper007]

Problem Statement

Customers can use it as a confidence check to choose correct tools.

Proposed Solution

Customers can use it as a confidence check to choose correct tools.

Use Case

Sometimes, the agent may pick up a wrong tools. Adding a built-in intent confidence, users can set up threshold to control the tools usage.

Alternatives Solutions

No response

Additional Context

No response

[Unshure]

Hi @leecooper007, Thanks for reaching out!

I'm curious if this is something you can accomplish with prompt engineering? Guiding an agent to choose the correct tool comes down to the context that is provided to the agent, through the system prompt and tool descriptions. You can use your agent and tweak these pieces of context to get better results from the tool choices of your agent. There are some useful links to prompt engineering guides at the bottom of our prompts documentation page: https://strandsagents.com/latest/documentation/docs/user-guide/concepts/agents/prompts/#prompt-engineering

[dbschmigelski]

Hi closing this out as is. As @Unshure said, we think this is something prompt engineering should handle now. That being said we would be interested to see an implementation using the new Hooks system or through a custom/extended Model provider.

I could see a Model provider returning Tool confidence level which can then be leveraged either in the Model or in some pre tool use hook.

Glad to discuss, but closing the issue out as this is not something we are eyeing to support natively at the moment.