When is it safe to run tool calls in parallel?

Safe: "get weather for 3 cities", "search 4 documentation sites", "fetch profile + orders + tickets for one customer."

Unsafe or needs care

• Dependencies: create_user then assign_role(user_id). The second needs the first's output.
• Conflicting writes: two edits to the same file or record. Race conditions or lost updates.
• Ordering semantics: "debit account A, credit account B" should be one transaction, not two parallel calls.
• Rate limits: 20 parallel calls can trigger 429s.

Executor design

• Tag tools as read_only vs mutating. Run read-only calls in parallel; serialise mutating calls, or lock per resource.
• Bound concurrency (semaphores) and handle partial failures: 3 of 4 calls succeed → return all results with clear errors for the failed one.
• Return results in a deterministic order, matched by call IDs.

Bonus: parallel sub-agents follow the same rule. Split work along independent lines, e.g. different research topics, not different parts of the same function.

Related

• Previous: Q29. Design long-term memory for an agent: what to store, when to write, how to retrieve, and how to forget.
• Next: Q31. Your agent worked great in the demo, but fails about 30% of the time in production. How do you approach it?