Local
The agent loop runs in the same process as the request handler.- Latency — lowest (no broker hops).
- Durability — none. If the process crashes mid-run, the run is lost.
- Concurrency — limited by the process.
- Best for — development, short interactive chats, low-risk operations.
Temporal
Each agent loop runs as a Temporal workflow; each IO operation (LLM call, tool call) is a Temporal activity.- Durability — workflows survive process restarts, replay deterministically from history.
- Retries — per-activity retry policy with exponential backoff.
- Long-running — workflows can run for hours or days; the agent can sleep on Delay nodes or wait for approvals without holding compute.
- Streaming — handled via a sidecar Redis stream so the client still sees token-by-token output.
- Best for — production agents, anything that calls expensive tools, anything you can’t afford to lose.
Restate
Each agent loop runs as a Restate workflow; each IO operation is a durable Restate step.- Durability — IO results are memoized by Restate. On recovery the workflow replays the recorded results instead of re-issuing the calls.
- Streaming — also Redis-sidecar based.
- Best for — teams already running Restate, agents with idempotency requirements around external side effects.