[Roadmap] Enable Parallel Tool Execution by Default

[VascoSch92]

Context

We have infrastructure for parallel tool calls (tool_concurrency_limit), but it defaults to 1 (sequential) to avoid breaking things. This issue tracks the investigation and work needed to safely increase the default.

Roadmap

The Roadmap is composed of 5 phases:

• [PTE - Phase 1] Analyze Tools for Parallel Safety Issues #2526
• [PTE - Phase 2] Design Solutions #2527
• [PTE - Phase 3] Benchmark Infrastructure #2528
• [PTE - Phase 4] Implement & Validate Solutions #2530
• [PTE - Phase 5] Gradual Rollout #2531

Open Questions

• What's the right default concurrency limit? (2? 4? 8?)
• Should we auto-detect conflicts or require tools to declare safety?
• How conservative should we be with unknown tools (MCP, custom)?
• How do sub-agents interact with parallelism? Inherit settings?

References

• Parallel Tool Calls Infrastructure PR: feat(sdk/agent): Parallel Tool Call Execution #2390
• Documentation PR: docs(sdk): Add guide for parallel tool execution docs#409
• Example: examples/01_standalone_sdk/45_parallel_tool_execution.py

[enyst]

What's the right default concurrency limit? (2? 4? 8?)

Just a thought. I feel this shouldn't have much of a limited answer. That is, I think maybe we can land on some definition which lets read-only tools be as many as the user configured, and a much higher default from us; higher than 8.