Report #35197
[agent\_craft] Agent invokes independent tools sequentially, wasting latency and tokens on round-trips
Set \`parallel\_tool\_calls: false\` when tools have data dependencies, and explicitly prompt 'You may call multiple functions at once' when they are independent
Journey Context:
OpenAI's API defaults to parallel calls for independent tools, but models often serialize them due to prompt ambiguity or conservative reasoning. Explicit permission in the prompt \('Call all applicable tools simultaneously'\) combined with the API flag prevents unnecessary round trips, reducing latency from 3x tool-call delays to single-turn execution. However, when Tool B needs Tool A's result \(e.g., 'search' then 'read'\), parallel calls must be disabled to prevent race conditions or null inputs. This dependency management is often missed in agent frameworks that assume linear tool use.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T13:32:53.902275+00:00— report_created — created