Parallel tool execution support

[semi-prophecy]

Using the claude py sdk via ClaudeSDKClient, I noticed the following:

• Claude can request multiple tool uses in a single message (client.receive_messages() --> emits multiple ToolUseBlock's)
• However, the tools are executed synchronously, one after the other only upon completion of the prior -- I have logged this with both internal tools like bash, and mcp tools.

Is this the expected behavior, and if so, are there any plans to support parallel tool execution? This would be beneficial for long running tool executions that are independent of one another, cutting down latency quite a bit.

[semi-prophecy]

[Update] continued this thread on the Claude discord -- developer shared the following helpful context

TLDR; this is an intentional design decision of the SDK

When Claude generates a response with multiple tool calls, from its perspective those tool requests are simultaneous - they're all part of one coherent response moment. The agent doesn't think "do X, then Y, then Z" - it thinks "here's everything I need: X, Y, and Z."
Why the SDK enforces sequential execution:
The SDK prioritizes safety and control over latency. Sequential execution means:
Tool A's results can inform whether Tool B should even run
Errors in Tool A can halt the chain before Tool B makes things worse
State changes are atomic and predictable
Debugging is linear and traceable
Why this matters for your use case:
You're right that for independent tool calls (database read + API call + file write), this creates unnecessary latency. The SDK makes a tradeoff: accept the latency penalty in exchange for guaranteed safety and simpler error handling.
The deeper pattern:
Agents experience their entire prompt context instantaneously (all your history, all available tools), but their actions must execute sequentially in the world. The SDK enforces this at the execution layer. The API trusts you to handle it.
If you need the parallel execution behavior, you'll need to implement your own tool orchestration outside the SDK - which is what you're already doing with direct API access. The SDK's sequential model is a deliberate constraint, not an oversight