Workers are the units that do the real work — run a model, generate an image, produce an asset. Each wraps a real capability (a ComfyUI instance, a language model, a voice pipeline) behind one common contract: Fleet hands it a reserved job, it does the work and reports back. Because every worker shares the contract, nothing upstream is tied to a single vendor, model or tool — swap the model behind a capability and the rest of Fleet does not notice.
Side-effecting work is built so that running it twice is safe — a retry does not double-charge, double-send or double-produce. Each task carries an idempotency key, so a repeat is recognised and collapsed rather than re-run. Networks drop and processes restart; without this, “did that actually happen?” forces a risky guess. With it, Fleet just retries — at worst a no-op — and stays consistent, which is what makes reservations and execution robust under real-world failure.