golem-parallel-workers-moonbit

Golem agents execute sequentially, so parallelism requires distributing work across multiple agent instances. Spawn child agents with AgentClient::scoped, trigger work fire-and-forget, and collect results by awaiting Golem promises.

Child agents have persistent identities and suit independent, stateless work, while @api.fork() clones the current agent with its state for ephemeral parallel execution. Forked agents are phantoms without persistent identity.

Create a promise per worker with @api.create_promise, pass each promise ID to a child agent, and have workers call @api.complete_promise with their result. The coordinator calls @api.await_promise for each ID to gather results.

Yes, use chunked fan-out by processing items in batches of a fixed size, such as five. Spawn and await each chunk fully before starting the next batch to bound the number of concurrent child agents.

Never have two agents await each other synchronously. Use fire-and-forget trigger calls to break cyclic waits, and synchronize through promises instead of blocking bidirectional calls.