Calling multiple tools to be run serially

[ryx2]

Description

Inspired by https://blog.cloudflare.com/code-mode/, I'm wondering if we can have a model tool output mode that do multiple tools serially.

The way I want it to be:

(1 LLM call)
output: tool1 -> tool2(tool1.result.value1) -> tool3(tool2.result.value2 + tool1.result.value1)

as opposed to the current iterative structure which requires far more agent loops

(1 LLM call)
output: tool1
(1 LLM call)
output: tool1 -> tool2(tool1.result.value1)
(1 LLM call)
output: tool1 -> tool2(tool1.result.value1) -> tool3(tool2.result.value2 + tool1.result.value1)

since this LLM is definitely smart enough to pipe the outputs into the next tool call, one shot.

References

No response

[DouweM]

We expect to ship full mcp-run-python based code mode before the end of the year as well: #2037.

It'd be interesting to build more light-weighed tool call chaining as well, possibly even leading up to that as it's cover some of the same ground. This should be implementable as a wrapper toolset that makes just a single "call_tool_chain" tool available and explains the available functions (including input and output) in its description.

We should be able to get quite far with https://pypi.org/project/simpleeval/ and validating type safety

[ryx2]

@DouweM my issue looks like a duplicate of that issue! I have some more comments, which I will leave in #2037 momentarily