FDE Times: Parallel tool calls — every failure still needs a tool_result, or the whole turn breaks

Customer-site tools are slow and flaky. FDE Times’ 5 November guide starts from a common failure mode: an agent issues several tool_use blocks in one turn, two succeed, a third times out, and handler code that returns only the successes breaks the whole turn. Anthropic’s docs (as cited) allow multiple tool_use blocks per turn; OpenAI likewise advises assuming zero, one, or many calls.

The build path is concrete: classify tools as read-only vs side-effecting before writing async code; wrap each call so timeouts and exceptions still produce a tool_result with is_error and an actionable message; run reads concurrently and writes in order, still emitting error results for skipped dependents; send every result back in one user message before any text. Anthropic’s Tool Runner is noted as an alternative when custom batching is unnecessary; OpenAI’s parallel_tool_calls=false can force serial calls.

FDEs are told to measure tool error rate, calls per task, and max call time on ticket-derived test cases—not demo scripts—so they can report a number to the customer’s ops team and tell a reliability story in interviews rather than “the agent worked in the demo.”