Temporal Serverless: Durable Execution Without Infrastructure
Cloud Platform EngineeringTemporal’s Durable Execution model is the infrastructure layer behind OpenAI’s production workflows, multi-cloud pipelines, and an expanding range of agentic AI systems. Its one persistent infrastructure cost has been the long-lived Worker process — a continuously-polling service that most teams have to design, scale, and operate. Replay 2026 changed that.
The Long-Lived Worker Problem
Temporal Workflows achieve durability by decoupling execution logic from the underlying compute. A Worker is the process that hosts Activities and polls Temporal Cloud for tasks to execute. The model is reliable: if a Worker dies mid-execution, another Worker can resume the Workflow from its last checkpoint without losing state. But the polling architecture assumes continuous running processes, which creates operational overhead that many teams find disproportionate to their workload profiles — particularly for agentic tasks that run in bursts and sit idle between invocations.
For teams building agentic workflows — where agent runs are triggered by events rather than continuous traffic — managing a fleet of always-on Workers to handle workloads that may be sparse is an infrastructure tax that doesn’t improve reliability. It just increases cost and operational surface area.
What Temporal Announced at Replay 2026
At Temporal Replay 2026 in San Francisco — attended by over 2,000 developers — CEO Samar Abbas and CTO Maxim Fateev introduced Serverless Workers for AWS Lambda, currently in pre-release. The mechanism: instead of a continuously-running process that polls for tasks, Temporal Cloud automatically invokes, scales, and gracefully shuts down Workers based on actual workload volume. When no work is queued, Workers scale to zero. When demand spikes, Workers spin up without developer intervention.
Deployment is a three-step process: upload Worker code to an AWS Lambda function, create an IAM role via CloudFormation for Temporal Cloud authentication, then register the function with Temporal. Supported SDKs at launch are Go, Python, and TypeScript, with additional languages and Google Cloud Run support on the roadmap. The AWS Lambda 15-minute execution timeout applies to individual Activities within a Workflow — long-running Workflows themselves are unaffected, as they span multiple Worker invocations and checkpoint state normally.
OpenAI VP Venkat Venkataramani, speaking at Replay, described Temporal’s role in production: “Temporal’s durable orchestration framework is critical to handling our massive scale, complex agentic workflows, infrastructure control plane, and data pipelines.” Temporal also announced Workflow Streams for real-time observability and Standalone Activities — currently in public preview for Go, Python, and .NET.
The So What
For teams already running Temporal for agentic workflow orchestration, Serverless Workers address the deployment model mismatch between event-driven agentic workloads and continuously-polling Worker infrastructure. Event-triggered agent tasks — the common pattern for multi-step research, code generation, and data processing workflows — are exactly what serverless compute is designed for. Teams should evaluate Serverless Workers when their Worker utilization is low relative to capacity (suggesting idle compute waste) or when infrastructure management is slowing agent deployment cycles. For teams evaluating Temporal for the first time, the serverless entry point eliminates the most common friction point — getting durable execution running without becoming a distributed systems operator first.
Photo by Julio Lopez on Pexels
Source: The New Stack
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →