[Reopen] AgentExecutor async path still routes through LLM.call

[suqinghen]

Summary

Issue #6683 was previously closed by inactivity, but the behavior is still present and impacts async pipelines.

In AgentExecutor.ainvoke()/run_async flow, LLM calls are routed through the shared synchronous path (get_llm_response()), even when the provider implements an async path (LLM.acall() / aget_llm_response()).

As a result, async workloads can be constrained by thread-pool capacity and lose the expected async execution benefits.

Current behavior

• AgentExecutor methods call_llm_and_parse and call_llm_native_tools call get_llm_response().
• get_llm_response() invokes LLM.call().
• LLM.call() uses sync execution path, requiring threadpool offload for providers that are actually async-capable.

Expected behavior

• Async execution paths should use aget_llm_response() and LLM.acall() when available.
• Preserve compatibility for custom LLMs that only implement sync behavior by falling back to asyncio.to_thread(...) only on NotImplementedError.

Suggested fix

1. Convert the two router methods to async and await aget_llm_response().
2. In aget_llm_response(), implement narrow fallback:
   • try async path first
   • on NotImplementedError, delegate to asyncio.to_thread(llm.call, ...) (or equivalent)
3. Keep sync behavior unchanged for non-async contexts.

Repro (minimal)

1. Create a custom LLM double that records calls for both call() and acall().
2. Attach it to an Agent and invoke await agent.kickoff_async("probe").
3. Validate that acall() is invoked when available.

If needed, I can provide a tiny regression test draft.

Context

This is a re-open-style report derived from #6683 and current source inspection.

[stud925]

Hi, I'd like to work on this. Happy to take it if unassigned.

[Vidit-Ostwal]

Sounds good, @stud925 assigning this to you.