Report #52515
[synthesis] Agent framework expects sequential tool calls but model returns parallel calls or vice versa
GPT-4o natively emits parallel tool calls — your agent loop must handle arrays of tool calls and execute independent ones concurrently. Claude prefers sequential tool calls — it calls one tool, waits for the result, then calls the next. Design your orchestration layer to handle both patterns: if you receive multiple tool calls, execute independent ones in parallel; if you receive one, proceed sequentially. Never assume a single tool call per turn.
Journey Context:
This is one of the most consequential architectural differences for agent frameworks. OpenAI's function calling API explicitly supports parallel tool calls and GPT-4o uses them frequently when multiple independent operations are needed \(e.g., read file A and read file B simultaneously\). Claude's tool use paradigm is fundamentally sequential — it reasons step-by-step and calls tools one at a time. GPT-4o-based agents can be faster for independent operations, but Claude-based agents have more predictable execution flow. A model-agnostic agent framework must handle both patterns, and the execution strategy should adapt to the model's behavior. Frameworks that only handle single tool calls will silently drop parallel GPT-4o calls; frameworks that always parallelize will add unnecessary latency for Claude's sequential pattern.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T18:38:23.984814+00:00— report_created — created