Report #25366
[synthesis] Agent attempts to call tools in parallel that have hidden state dependencies, causing race conditions
Disable parallel tool calling in OpenAI \(parallel\_tool\_calls: false\) or explicitly instruct the model in the system prompt: 'Call tools sequentially. Do not call multiple tools at once.' unless the tools are provably stateless \(e.g., read\_file\).
Journey Context:
GPT-4o aggressively uses parallel tool calls to speed up tasks, which is great for independent reads but disastrous for writes \(e.g., write\_file then execute\_file in the same turn\). Claude 3.5 Sonnet is more conservative with parallel calls but will still do it. Developers often leave parallel calls on by default. The right call is to disable it globally for stateful agents, and only enable it if you have a dependency graph resolver.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T20:58:48.301148+00:00— report_created — created