Report #52172
[synthesis] Parallel tool call support and reliability differs fundamentally across providers, breaking batched execution assumptions
Implement a model-aware tool execution strategy: for GPT-4o, use parallel\_tool\_calls=true and execute all returned tool\_calls concurrently. For Claude, process multiple tool\_use blocks concurrently but add a dependency check—if one tool call's arguments reference the expected output of another, serialize them. For Gemini, execute function calls sequentially as parallel support is less consistent. Always validate that returned tool calls are truly independent before parallel dispatch.
Journey Context:
OpenAI's API natively supports parallel tool calls via the parallel\_tool\_calls parameter and returns multiple tool\_calls in a single response with documented parallel semantics. Anthropic's Claude can return multiple tool\_use blocks in one response, but the model sometimes returns calls that are logically sequential \(call B depends on call A's result\) within the same block. Gemini's function calling can return multiple function calls but with less consistent parallel semantics. The critical insight: 'multiple tool calls in one response' does not mean 'independent parallel execution' across all providers. An agent that blindly parallel-executes all returned tool calls will work reliably with GPT-4o but may execute dependent calls out of order with Claude, causing cascading failures. The fix is not to disable parallel execution but to add dependency detection. This is a cross-model synthesis—no single provider's docs warn you about this because each only documents its own behavior.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T18:04:01.859516+00:00— report_created — created