`ParallelAgent` executes sub-agents sequentially instead of concurrently

[fedorovychh]

The Bug
Despite its name, ParallelAgent executes its sub-agents sequentially on a single thread rather than in parallel. This defeats the purpose of using a parallel agent, as total execution time equals the sum of all individual agent execution times rather than the longest single execution time.

To Reproduce

1. Configure a ParallelAgent with multiple sub-agents.
2. Trigger the agent execution using Runner.runAsync(...) and subscribe to the resulting Flowable.
3. Observe the logs or execution times: the sub-agents process their requests one after the other, not simultaneously.

Root Cause Analysis
The bug is located in ParallelAgent.java inside the runAsyncImpl method.

• The method maps over currentSubAgents and calls runAsync on each to create a list of Flowable<Event> streams.
• It then passes this list directly into Flowable.merge(...).
• Because it does not explicitly apply a concurrent scheduler (such as .subscribeOn(Schedulers.io())) to the individual sub-agent flows before merging them, RxJava defaults to executing the merged "cold" streams sequentially on the caller's thread.

Expected behavior
ParallelAgent should internally handle thread scheduling for its sub-agents (e.g., dispatching them to an I/O thread pool before the merge step) so that Flowable.merge() can actually process their streams concurrently.

[hemasekhar-p]

Hi @fedorovychh, Thank you for reporting this issue,

I’ve verified this issue in the latest ADK (v0.6.0). By default, ParallelAgent executes sub-agents sequentially on the subscribing thread because runAsync() returns cold RxJava Flowables and no scheduler is applied. To enable true parallel execution, each sub-agent should be wrapped with subscribeOn(Schedulers.io()) before merging. With this change, sub-agents run concurrently on separate threads, and total execution time becomes equal to the longest sub-agent instead of the sum. I recommend updating ParallelAgent to apply this scheduler by default to match the intended parallel semantics.