Please read this first

• Have you read the docs? Agents SDK docs
• Have you searched for related issues? Others may have had similar requests

Describe the feature

I've seen some posts recently around tool reliability, handing things off between agents, and running into duplicate tool names across MCP servers. Has the OpenAI team considered adding an optional intent classifier for agents needing lots of tools? E.g., only providing the agent with tools that are relevant to the prompt?

For instance, there's no need to send 20-30 tools from attached MCP servers for simple prompts like "Hello" or, say, provide tools from Supabase's MCP server if I'm trying to have my agent do something through GitHub. The SDK currently (like all agent frameworks) loads and passes all attached tools on each prompt, which can bloat the context window or confuse the model.

We're working on an API to solve this at Proxy, and I think as an optional plugin it could be really helpful for use cases with many tools. I did see that the team is discussing how best to add third-party integrations, though, and don't want to clutter the repo if this isn't something that the OpenAI team doesn't want added just yet.

Happy to create a PR with a minimal example if it's something the OpenAI team would be open to!