golem-parallel-workers-scala

Golem agents execute sequentially, so parallelism is achieved by spawning multiple child agent instances via the codegen-generated XClient.get(id) and collecting their Futures with Future.sequence. Alternatively, use HostApi.fork() to clone the current agent for lightweight parallel execution.

Child agents have persistent identity and suit independent, stateless work, while fork() clones the current agent with its state into an ephemeral phantom instance. Use child agents when workers need their own ID, and fork() when workers must inherit the parent's current state.

Use Future.sequence to combine the Futures returned by child agent calls into a single list of results. For fire-and-forget dispatch, create a Golem promise per child with HostApi.createPromise() and await each with HostApi.awaitPromise or awaitPromiseJson.

Wrap each child call with .map(Right(_)).recover to capture failures as Left values, then partition the sequenced results into successes and failures. This lets the coordinator return partial results instead of failing the entire fan-out.

Yes, two agents synchronously awaiting each other can deadlock. Break cycles by using .trigger for fire-and-forget invocation combined with Golem promises, so agents suspend on promise completion rather than blocking on direct RPC.