Why some calls run together
When the model requests multiple tools in one turn, the agent checks their categories. Reads and subagent tasks can run concurrently because they are intended to gather information without changing the target. This can shorten research tasks that inspect independent files or sources.
If a group includes an action that may ask for permission, the agent executes those calls sequentially. That gives each proposed change a clear approval point and avoids presenting several permission requests at once. The scheduling behavior does not replace reviewing what a tool will do; check the action and its target before allowing a change.
Plan independent work clearly
Parallel reads are most useful when each call can proceed without the result of another. If one operation depends on a value discovered by a previous operation, let the agent finish that step first and use its result in a later call. This keeps dependencies explicit and avoids guesses based on unfinished work.
Concurrency also does not guarantee that external services respond in a particular order. For a task that writes shared data, asks for approval, or relies on a precise sequence, describe the required order directly. The agent's report should distinguish completed calls from failures and permission refusals.