diff --git a/docs/best-practices/task-orchestration.mdx b/docs/best-practices/task-orchestration.mdx
index 28d0e35d5..52b2abf43 100644
--- a/docs/best-practices/task-orchestration.mdx
+++ b/docs/best-practices/task-orchestration.mdx
@@ -238,6 +238,72 @@ analysis_tasks = {
 }
 ```
 
+### Async Loop Expansion
+
+A `task_type="loop"` start task expands into one subtask per CSV/text row on the async path as well as the sync path.
+
+<Note>
+**Async parity ([PR #3326](https://github.com/MervinPraison/PraisonAI/pull/3326)):** A `task_type="loop"` start task now pre-expands one subtask per input-file row on the **async** path — `await process.aworkflow()` and `process.workflow()` produce the same subtask set. Before this fix, the async engine skipped pre-expansion and ran the loop task once as an ordinary task, so a 500-row input file processed only the first row.
+</Note>
+
+```python
+import asyncio
+from praisonaiagents import Agent, Task, Process
+
+analyzer = Agent(name="Analyzer", instructions="Analyze one row at a time")
+
+loop_task = Task(
+    name="analyze_row",
+    description="Analyze row {row}",
+    agent=analyzer,
+    task_type="loop",
+    input_file="rows.csv",   # one Task per row, on the async path too
+    is_start=True,
+)
+
+process = Process(tasks={"analyze_row": loop_task}, agents=[analyzer])
+
+async def main():
+    async for task_id in process.aworkflow():
+        print(f"Completed: {task_id}")
+
+asyncio.run(main())
+```
+
+For an N-row CSV, the async iteration yields N subtasks (`analyze_row_1`, `analyze_row_2`, …). The parent loop task is marked `completed` after pre-expansion; the subtasks carry the actual work.
+
+```mermaid
+graph LR
+    subgraph "Loop-task expansion"
+        Csv[📄 rows.csv] --> PreExp[🧩 pre-expansion]
+        PreExp --> S1[loop_start_1]
+        PreExp --> S2[loop_start_2]
+        PreExp --> S3[loop_start_3]
+    end
+
+    subgraph "Engines (same output)"
+        Sync[⚙️ workflow]
+        Async[⚡ aworkflow]
+    end
+
+    S1 --> Sync
+    S1 --> Async
+    S2 --> Sync
+    S2 --> Async
+    S3 --> Sync
+    S3 --> Async
+
+    classDef input fill:#8B0000,stroke:#7C90A0,color:#fff
+    classDef process fill:#189AB4,stroke:#7C90A0,color:#fff
+    classDef sub fill:#F59E0B,stroke:#7C90A0,color:#fff
+    classDef engine fill:#10B981,stroke:#7C90A0,color:#fff
+
+    class Csv input
+    class PreExp process
+    class S1,S2,S3 sub
+    class Sync,Async engine
+```
+
 ### The Retry Pattern
 
 Implement robust retry logic:
diff --git a/docs/features/thread-safety.mdx b/docs/features/thread-safety.mdx
index 7e1a8ef3c..b96a60a4b 100644
--- a/docs/features/thread-safety.mdx
+++ b/docs/features/thread-safety.mdx
@@ -453,6 +453,7 @@ PraisonAI provides comprehensive thread-safety for HTTP server deployment:
 - If two launch calls use the same path on the same port, the second gets an auto-suffixed path (`/path_abc123`) and a warning is logged.
 - Server readiness is signalled deterministically (no fixed sleep); `.launch()` returns only after the port is accepting connections. The wait defaults to **5 seconds** and is configurable via the `PRAISONAI_SERVER_READY_TIMEOUT` environment variable. If the server doesn't become ready in time, `.launch()` still returns and a warning is logged — check server logs for startup errors.
 - `aworkflow()` state lock is created inside the running async context, so workflows remain stable when invoked under pytest-asyncio or when nested inside another loop.
+- Per-turn tool tracking (`_turn_tools_used`) is now protected by the same lock that guards `chat_history`, so concurrent `chat()` / `achat()` turns on one `Agent` no longer corrupt hook or self-improve tool data. No new public API — the buffer stays private; the read/append helpers were made async-safe internally.
 
 ```python
 import threading