Hi everyone,
I'm trying to understand whether I'm hitting a platform limitation or if there is another issue involved.
I have a Power Automate flow that uses a For Each loop with concurrency enabled. The goal is to execute multiple Copilot Studio agent requests in parallel in order to reduce document processing and information retrieval time.
With a concurrency setting of 50, the flow launches 50 simultaneous "Execute agent and wait" actions against the same Copilot Studio agent.
The approach works very well from a performance perspective, but I start seeing many failures when the level of parallelization becomes high.
The failures are not always the same. In many cases, the agent's SharePoint search seems to stop working, and sometimes the action returns the following error:
What makes troubleshooting difficult is that the behavior appears quite random:
- I never get exactly the same number of failures.
- Sometimes more than half of the 50 requests fail.
- Sometimes fewer requests fail.
- The failures seem to occur on different iterations each time.
However, when I reduce the concurrency to fewer than 10 simultaneous agent calls, the agent works very well and in most cases I observe no failures at all.
My flow structure is relatively simple:
The relevant action is:
Additional context:
- The agent performs SharePoint document searches.
- The same agent is used for all requests.
- Each request is independent.
- Lower concurrency levels (<10) are reliable.
- High concurrency levels (50) produce intermittent failures and timeout errors.
My questions are:
- Are there documented limits on the number of simultaneous calls to a Copilot Studio agent from Power Automate?
- Are there throttling limits related to agent execution, SharePoint search, or the ExecuteCopilotAsyncV2 action?
- Could the
AIModelActionRequestTimeouterror indicate backend throttling rather than an actual timeout within the agent? - Has anyone successfully executed dozens of agent calls in parallel, and if so, what concurrency levels proved reliable?
Any information about platform limits, throttling mechanisms, or recommended architecture patterns for large-scale parallel execution would be greatly appreciated.
Thanks!

Report
All responses (
Answers (