How to dynamically add/remove tools in a `tool_use_behavior="run_llm_again"`-loop

[Afstkla]

Please read this first

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

Question

I'm building an agent that's often running multiple tool calls sequentially, and sometimes tool A will 'unlock' tools B, C, and D. But I so far haven't found a way to 'inject' those during an automated run_llm_again loop. And I'm really dreading writing all the logic required to manually handle the run loop.

I've attempted using the Hooks to dynamically add new tools, I've been trying to work with various different Agents as tools, where the tools of those Agents are added dynamically through the hooks. The only thing that's semi worked for me is to use the handoff feature, but then I can't give control back to the 'main' agent.

My agent's flow will normally be like:

1. User asks for information from Gmail
2. Agent Runner starts
3. Agent detects that there's no Gmail connection set up
4. Agent will call the "initiate gmail connection" tool
5. Response from the tool 'unlocks' all Gmail actions
6. Agent calls the 'get latest emails' tool
7. Agent streams the response
8. Agent Runner exits

But this is a workflow that I can't currently figure out. Is there a way?

[rm-openai]

This is a good question. I'll have to think of a long term good answer, but the short term hack is: just add the tools in the initiate_gmail_connection tool call.

my_agent = Agent(instructions=[...])

@function_tool
def initiate_gmail_connection():
   my_agent.tools.extend([send_gmail, read_gmail, ...])

[Afstkla]

Thanks @rm-openai , unfortunately, it seems that extending the tools isn't enough reason to refresh the current_span / agent_span in the _run_streamed_impl (

openai-agents-python/src/agents/run.py

Line 518 in 6e078bf

if current_span is None:

), and therefore extending the tools doesn't work within the run_llm_again loop (or I tested it wrong, ofc also possible).

Any other ideas?

[rm-openai]

Ah you're totally right. Let me fix this - we shouldn't be caching the tools list.

[rm-openai]

@Afstkla this is now fixed. I'd recommend using is_enabled on the tool, which lets you specify a function to dynamically enable/disable behavior.