LLMs, GenAI, RAG, agents, evaluation · insight-1568
Run independent read-only tool calls in parallel and serialise writes
- Evidence type
- documented fact
- Difficulty
- easy
- Roles
- ML engineer, Applied AI
- Industries
- aviation
Summary: situation, action, reason, boundary
Situation
An agent needs a flight status, a fare rule and a loyalty record before answering, and issues the lookups one after another.
Action
Execute the calls the model emits in one turn concurrently when they are independent and read-only, run side-effecting or order-dependent ones sequentially, and always return a result for every call, marking skipped ones as errors.
Reason
The provider leaves execution order to the application and notes that independent read-only operations are usually safe in parallel while side effects may need sequencing.
Boundary
Agents that emit several tool calls per turn.
Symptom or risk
Latency is the sum of the calls, and the user waits, while naive parallelism of writes can produce conflicting side effects.
Exceptions
Rate-limited backends may need a concurrency cap.
Trade-offs
Concurrency adds error handling, and partial failures need clear reporting.
Pitfalls
Omitting a result for a skipped call, which breaks the message format.
Lessons and concepts
Claims and their sources
Anthropic's documentation says the application decides how to execute multiple tool calls, that independent read-only operations are usually safe in parallel, and that a result must be returned for every call, with unrun calls marked as errors.