Glossary
parallel tool calls
When a model emits several independent tool calls in a single turn so the host can run them concurrently, instead of one call per round-trip.
A single turn that contains more than one tool call. When the tasks do not depend on each other (look up the weather in three cities, read four files), a model that supports parallel calls emits them together and the host runs them at once, then returns all the results before the model continues. The alternative is one call per tool-call loopagentsThe host-orchestrated cycle behind tool use: the model emits a tool call, the host (not the model) executes it, and the result is fed back for the next turn. Open full entry iteration, which serializes work that did not need to be serial.
The payoff is latencycomputeThe time from request submission to response completion, broken down for LLMs into time-to-first-token and time-per-output-token, the user-facing speed metric. Open full entry . Each round-trip through the model costs a full generation, so collapsing four sequential calls into one parallel batch removes three of those generations from the critical path. The cost is that the model has to recognize independence correctly; calls that actually depend on one another must stay sequential, and a model that parallelizes a dependent chain produces wrong results.
Support and the exact request flag vary by model and provider, so the capability is best described generically: the model can return a list of calls in one turn, and the host fans them out. It pairs naturally with function callingagentsA pattern where a model emits a structured call (function name plus arguments), the runtime executes it, and the result returns as input on the model's next turn. Open full entry and is one of the levers that separates a snappy agent from a sluggish one.