@calm-artist-7206 Hey, when you have a moment. I answered your agent questions 🙂 https://github.com/OpenHands/software-agent-sdk/pull/2390

Thank you! I don’t really see issues with the PR, it’s great to see and great to try yummy parallel tool calls! I would add, just as a tiny aside. I feel that my slow brain, skimming through another layer of abstraction with multiple lambdas and callables, is losing track. 😅 Maybe that's ok! Hey, it's not the best brain ever but oh well, it's the one I haz. lolsob Just a thought: I think the PR makes this extra layer of abstraction on top of execution layer precisely to keep more complexity out of the agent. But maybe the execution layer itself could be out. Maybe worth making a mental note like, e.g. next time we're looking or when we're breathing a bit, we could put on the table some re-design of the execution layer out of the agent. Then idk, orchestrating parallel tool calls could fall into place more easily maybe, though I could be wrong

One thing that would be nice, would be maybe an integration test (a real-world-like one) or trying somehow to make sure that one or more of the existing integration tests uses them

I don't think it is necessary. I mean, it is necessary just to check that LLM are emitting multiple tool calls in the way we are expecting. But I think this is already the case right?

Is that also for e.g. multiple

bash

calls?

Why should not be the case? Can you give me a concrete example?

Not really. 😅 I simply don’t know if the execution in the terminal session will work ok in parallel, or it has gotchas, Xingyao’s expertise needed

I'm creating an example file where we can check that everything works 🙂