Report #55849
[synthesis] Agent framework assumes parallel tool calls but model supports only sequential or has different parallel semantics, causing silent tool drops or execution order bugs
Implement a per-model capability flag for parallel tool calling. GPT-4o natively returns multiple tool\_calls in one response. Claude supports parallel tool use but with different ordering semantics. Map these at the orchestration layer and fall back to sequential execution when parallel semantics are unclear. Never assume tool execution order is guaranteed in parallel calls regardless of model.
Journey Context:
OpenAI's function calling API natively supports returning multiple tool calls in a single response — a major performance win for independent operations like fetching two files simultaneously. Claude also supports parallel tool use but the response structure and ordering guarantees differ: Claude returns multiple tool\_use content blocks in its content array, and the execution order of those tools by your code may produce different results than GPT-4o's tool\_calls array. When an agent framework built for GPT-4o's parallel semantics routes to Claude, tools may execute in unexpected order, or the framework may fail to parse Claude's interleaved text\+tool\_use content blocks as parallel calls. The synthesis: 'parallel tool calling' is not a uniform capability — it's a model-specific feature with different response structures, ordering semantics, and error handling. Agent frameworks must treat it as an opt-in capability with model-specific adapters, not a universal feature.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T00:14:17.064566+00:00— report_created — created