Deferred Tool Calls

[felixs8696]

Describe the feature

Sometimes tool lists get large and are unnecessary for all queries, incurring unnecessary cost, but more importantly unnecessary noise to the Agent. Instead, it's sometime better for tools to be dynamically loaded in as decided by the agent. This way the agent is aware of all of the available possible groups of tools, but only looks them up and their definitions when it needs them.

For example, if I have a Google Drive set of tools, and a linear set of tools, and a Slack set of tools, if a user just asks me to look up stuff in Google Drive, the tools available to the agent at the time should only be the Google Drive tools. But if the user says, "I want to look at something in Google Drive and then send it on Slack," then both sets of tools should be made available to the agent dynamically. Is this possible with the OpenAI Agents SDK? How would you recommend we proceed even if it's not?

https://www.anthropic.com/engineering/advanced-tool-use

[seratch]

Thanks for asking the question. We provide tool filtering for your need. https://openai.github.io/openai-agents-python/mcp/#tool-filtering

[felixs8696]

Thanks @seratch I'll experiment with this. Does this only work with an MCP configuration, not with a raw tool list?

Also when is this tool context filter hook called?

I have a use case where the agent can dynamically decide to enable tools during its Runner.run loop. If the tool content t filter isn't implements like a lifecycle hook that reeevalautes every turn then it won't work for my use case