Temporal

The SDK defines the API surface: Workflows (deterministic orchestration logic), Activities (non-deterministic side effects), Signals (external input to running workflows), Queries (read-only access to workflow state), and Updates (synchronized mutations). Workflow execution starts via the SDK client: client.workflow.start(workflowFunction, {args, taskQueue, workflowId}). The taskQueue connects workflows to workers. Workers poll task queues and execute workflow/activity code. The HTTP API at the Temporal Server provides workflow listing, description, termination, and signal delivery for administrative operations. Search attributes enable querying workflows by custom metadata. Visibility APIs list and filter running/completed workflows. The programming model is powerful: workflows can sleep for days, wait for signals, spawn child workflows, and compose activities — all with durable execution guarantees.

Activity retries are automatic with configurable retry policy (initial interval, backoff coefficient, max interval, max attempts). Workflow task retries handle transient failures in the workflow execution itself. The durable execution model means workflow progress is never lost: if a worker crashes mid-activity, the activity retries on a healthy worker. Heartbeating enables long-running activities to report progress and detect cancellation. Workflow timeouts (execution timeout, run timeout, task timeout) provide multiple layers of deadline enforcement. Dead letter queues aren't needed — failed activities retry according to policy, and permanently failed workflows can be reset. The main reliability consideration: Temporal Server itself must be highly available. Temporal Cloud handles this; self-hosted requires proper cluster configuration. For agents, the durable execution guarantee is the killer feature — mission-critical workflows never silently drop.

Temporal provides durable execution — workflows that survive process crashes, server restarts, and infrastructure failures. The programming model writes workflows as regular code (Go, Java, TypeScript, Python, .NET, PHP) with the Temporal SDK handling durability transparently. Activities (side effects like API calls) are retried automatically. Workflow state is persisted and resumed after any interruption. For agents building mission-critical workflows — payment processing, order fulfillment, long-running approval chains — Temporal provides guarantees that simpler job platforms don't: exactly-once workflow execution, automatic compensation (saga patterns), and versioning for live workflow updates. The trade-off is complexity: Temporal's programming model and operational requirements are more involved than Inngest or Trigger.dev. Temporal Cloud provides managed infrastructure; self-hosted requires running the Temporal server cluster.

Temporal Cloud uses namespace-based isolation with mTLS certificate authentication. Each namespace has its own certificate pair — workers and clients must present valid certificates to connect. API keys are available for programmatic access. Self-hosted Temporal supports configurable auth including mTLS, JWT, and custom authorizer plugins. Namespace isolation means different teams or environments can operate independently. The security model is enterprise-grade: mutual TLS is the default for Temporal Cloud, not an add-on. For agents, the certificate-based auth is more secure but more complex to configure than API key-based systems. Temporal Cloud's CLI handles certificate management. Self-hosted deployments have full control over the auth infrastructure.

Documentation at docs.temporal.io is extensive — covering concepts, SDK references for all languages, operational guides, and Temporal Cloud administration. The conceptual documentation is deep: it explains the event-sourcing execution model, deterministic constraints, versioning strategy, and failure handling in detail. SDK documentation varies by language — Go and Java are the most mature, TypeScript and Python are catching up. The learning curve is the steepest among background job platforms, and the documentation reflects this: there's a lot to learn. Temporal University provides structured courses. Community Slack is highly active with Temporal engineers participating. The documentation's main challenge: the conceptual depth can be overwhelming for newcomers. For agents already familiar with distributed systems, the documentation is thorough and authoritative.