LLMOrchestrator and its agents do not distinctively process separate workflow instances

[KaiWalter]

I am testing LLMOrchestrator with 2 agents with this repository.
This post explains the background and what I want to achieve.

Structure of LLM orchestrated flow:

                  Agent Task Planner
                /
"Intent" Orchestrator
                \
                  Agent Office Automation

don't get stuck on the naming, it is working in progress

The idea is, that the orchestrator gets a task from a preceeding other workflow over pubsub. The orchestrator with the agents extracts the intent of the task passed and uses tools to act on that intent.

As long as I only pass in on task: OK

As soon as I pass in 2 tasks in short sequence, in my case sending in 2 voice transcriptions to extract intent from ...

$ la .data/local_voice_inbox
total 124K
-rw-rw-r-- 1 kai kai 75K Aug 24 15:40 test20250824T1540480000.mp3
-rw-rw-r-- 1 kai kai 45K Aug 24 15:40 test20250824T1540490000.mp3

... steps in one of the workflow instances get not cleanly processed and in my sample one instance is trying to handle both files.

I attached state files

IntentOrchestrator_state.json

TaskPlanner_state.json

OfficeAutomation_state.json

Is there something I need to adjust to get a clean separation and parallel processing?

[bibryam]

Hi Kai, I love the example.
But I don't see LLMOrchestrator in the repo. Is it not there yet?

[sicoyle]

closing as this should be resolved now 🙌

[KaiWalter]

confirmed - my problem is fixed now

however, when upgrading to 0.10.5 I run into a different problem: #337