September 2026 Summaries

Temporal announced general availability for its Rust SDK and Standalone Activities, alongside Worker Controller autoscaling support for Kubernetes HPA and KEDA, enabling independent scaling of active Worker versions. The company also introduced a public-preview Temporal Cloud CLI extension with browser-based authentication, eliminated Temporal Cloud’s monthly base fee in favor of pay-as-you-go pricing, and offered introductory credits for new users. The update highlights design patterns, tutorials, and community examples, including an order-processing Workflow built with Encore and a field-operations case study from Xgrid involving intermittent connectivity. It also covers community guidance on replication warnings and limiting in-flight asynchronous jobs, promotes upcoming global events and webinars, and notes that Temporal raised $550 million at a $12.55 billion valuation.

Temporal’s Rust SDK is now generally available, enabling developers to create durable Workflows and Activities, operate Workers, and interact with Temporal through native Rust APIs integrated with Tokio. The SDK manages long-running application coordination by recording Workflow progress in Event History, allowing another Worker to replay that history and resume execution after failures. Rust structs, async functions, procedural macros, and typed method references support strongly typed Workflow and Activity definitions, while Activities can perform ordinary asynchronous I/O and side effects but should be idempotent because they may be retried. Workflow code must remain deterministic during replay, using Temporal-provided timers, workflow time, and concurrency primitives instead of direct system time, I/O, or nondeterministic Tokio operations; a built-in detector identifies many unsupported asynchronous operations. The SDK includes testing support, replay-based compatibility checks for Workflow updates, and capabilities such as Signals, Queries, Updates, retries, cancellation, Child Workflows, and tracing, supporting use cases including payments, provisioning, data pipelines, automation, approvals, and AI agents on Temporal Cloud or self-hosted deployments.

Temporal has made Standalone Activities generally available across Go, Python, Java, TypeScript, .NET, and Ruby, positioning them as a durable alternative to basic background-job queues that require teams to build separate tracking, retry, monitoring, and recovery systems. Each submitted job creates a persistent Activity Execution record that tracks its status, retries, errors, results, and lifecycle through Worker crashes, deployments, and timeouts, while supporting delayed starts, heartbeats and checkpoints for long-running work, search attributes, and individual operational controls. Because execution is at least once, developers remain responsible for making external side effects idempotent, such as using stable webhook update IDs for deduplication. Standalone Activities also offer priority and optional weighted fairness to manage shared Worker capacity, and can later be reused as steps within more complex Temporal Workflows without rewriting the Activity code. Available in Temporal Cloud and self-hosted deployments, the feature supports incremental adoption for use cases including webhooks, payouts, compliance checks, provisioning, notifications, and agent actions; Coinbase reports migrating part of its high-volume background-job traffic to Standalone Activities without downtime or job loss.

Temporal has introduced a prerelease integration that allows Amazon Bedrock AgentCore Runtime to serve as a compute provider for Temporal Serverless Workers, combining elastic AWS-hosted agent infrastructure with Temporal’s durable workflow execution. Developers can deploy agent code, such as the provided Python Strands sample, to AgentCore Runtime and configure it through a versioned Temporal Worker Deployment, enabling controlled rollouts and automatic scaling without maintaining continuously running Worker fleets. The approach is designed for bursty, long-running agent workloads, with Temporal preserving workflow state, retries, timers, signals, and recovery from failures while AgentCore provides managed compute, identity, tool access, policy, memory, and observability services. In the reference architecture, Temporal Workflows coordinate agent loops and long waits, while model calls and tool operations run as Activities with explicit retry and failure controls; AgentCore Runtime sessions host the Workers that poll Temporal task queues. The integration also supports efficient use of Runtime sessions during active model or tool calls, releasing capacity after work subsides, while maintaining durable application progress even when individual Workers are replaced.

KERNEL uses Temporal to manage authentication as a long-running connection lifecycle rather than a one-time login request, orchestrating nearly 80,000 successful sessions in 30 days, including over 5,000 that required human input or external approval. Each connection is represented by a durable Go-based parent Workflow responsible for configuration, health checks, reauthentication, cancellation, and failure handling, while individual login attempts run as child Workflows with separate browser instances, deadlines, and cleanup processes. Human actions such as MFA codes, account selection, push approvals, and CAPTCHA resolution are modeled as durable pause-and-resume states that survive Worker restarts without holding open HTTP requests or agent processes. TypeScript Workers perform model-driven browser-agent tasks through Activities and reattach to browser sessions when needed, while Go retains durable state and keeps credentials out of workflow event histories through short-lived leases. The system also performs recurring authentication checks, distinguishes inconclusive website failures from actual logout events, automatically reauthenticates when possible, and uses circuit breakers to avoid repeated failures or account lockouts. The approach emphasizes selecting the durable business lifecycle as the Workflow boundary, separating nondeterministic browser and model operations from persistent orchestration, and treating continued authenticated access as an ongoing product responsibility.

AI agent safety requires not only auditing actions after they occur but also enforcing clear limits on what agents may do before execution. While frontier AI labs must decide which capabilities to release, deploying companies are responsible for determining the authority granted to agents, including applicable policies, approval requirements, and credential access. These controls must remain reliable through production failures, timeouts, and retries so that an interrupted or rejected action cannot bypass established limits. Using a travel-booking example, the author argues that retries should preserve an approved $900 fare rather than revert to an unapproved $2,000 option. The author presents Durable Execution as a way to preserve approvals, rejections, and outcomes across failures, and describes Temporal Agent Harness as an execution-layer control point that works alongside existing agent development tools while leaving companies responsible for their own policies, identities, permissions, and security systems.

Temporal’s leadership describes growing demand for reliable AI-enabled applications as companies contend with increasingly complex legacy systems, positioning durable execution as essential infrastructure for production AI agents. Over the past year, the company expanded its leadership team, reached 570 employees across four continents, and reported 200% annualized revenue growth, net dollar retention above 200%, 1.9 trillion monthly Actions processed in August, 4,300 paying customers, and more than 43 million open-source installs. Customers including Snap, JPMorgan Chase, and OpenAI use Temporal for large-scale and regulated workloads, while the company also expanded its public presence through a partnership with Crystal Palace and its fifth Replay conference. Temporal released general-availability integration with the OpenAI Agents SDK and an integration with Vercel’s AI SDK, aiming to provide durable execution for agents that perform multi-step business actions. Looking ahead, the company plans further enterprise investment, global expansion, and platform development as it seeks to serve as a reliability layer for agentic AI systems.

Temporal is introducing a revised Temporal Cloud pricing structure that will apply to new customers in January 2025 and begin migrating existing pay-as-you-go customers in February 2025, while customers with existing credits retain current pricing until those credits are used. The new model offers four plans—Essentials, Business, Enterprise, and Mission Critical—with bundled Actions, active and retained storage, support levels, and selected features such as SSO for higher-tier plans. Pricing combines a base monthly or annual plan fee with usage charges beyond included allocations, and some plans apply a percentage-based fee as usage scales. Volume discounts will become available to all accounts starting at 5 million Actions and will aggregate usage across namespaces, while the initial per-million Action price rises to $50 and retained-storage pricing increases, with active-storage pricing unchanged. Temporal is also adding one-, two-, and three-year minimum-spend commitments that provide further discounts and fixed Action rates without punitive overages. The company states that the changes are intended to simplify billing, improve predictability, and expand access to volume discounts, though cost effects will vary by customer usage and support needs.

Temporal announced a $550 million Series E funding round at a $12.55 billion valuation, co-led by Lightspeed alongside several new and returning investors, to expand its platform for reliable long-running software workflows. The company positions its Durable Execution technology as infrastructure for AI agents and enterprise applications that must coordinate across complex systems for days or months, preserving state and recovering automatically from outages or failures. Founded by Samar Abbas and Maxim Fateev, whose prior work included Amazon SWF, Microsoft Durable Task Framework, and Uber Cadence, Temporal has focused on distributed workflow orchestration since before the recent AI boom. It reports annualized revenue growth of more than 200% year over year, net dollar retention above 200%, 1.9 trillion billable actions processed in August, more than 43 million open-source installations, and over 4,300 paying customers including OpenAI, Snap, NVIDIA, and JPMorgan Chase. The funding will support global expansion, core platform development, and enhanced reliability and security capabilities for enterprise users.

Temporal Cloud has introduced a pay-as-you-go offering with Developer Support aimed at AI and application teams that want to experiment without committing to a monthly base fee or forecasting usage. The service preserves workflow execution state through failures and supports work across models, tools, systems, and time, allowing projects to scale into production on the same platform. Pricing includes $50 per million Actions, $0.042 per GB-hour for Active Storage, $0.00105 per GB-hour for Retained Storage, and Developer Support at 10% of usage with a one-business-day target for P0 issues; SAML SSO is included, while SCIM is optional. The plan has no feature gates, offers $150 in credits for 90 days to new users, and lets customers later select faster support or discounted committed pricing without changing product capabilities.

Temporal Cloud Serverless Workers on AWS Lambda provide an on-demand alternative to long-lived Kubernetes Workers for bursty, unpredictable, or low-volume workloads, while dedicated Workers remain preferable for steady, high-throughput traffic. Serverless Workers use the same Temporal SDK, Workflow, and Activity code but are invoked by Temporal’s Worker Controller Instance when Task Queue demand or backlog requires capacity, then shut down after work drains; each invocation creates a new client connection and Lambda limits individual Activity execution to at most 15 minutes, though Workflows can span unlimited invocations. Migrating involves adding a Lambda handler, configuring Worker Versioning with PINNED or AUTO_UPGRADE behavior, carefully coordinating graceful shutdown, deadline buffer, and Lambda timeout settings for long Activities, packaging Linux-compatible dependencies within Lambda’s size limits, and creating separate IAM roles for Lambda execution and Temporal Cloud invocation. The process also includes deploying and versioning the function, registering and activating its Worker Deployment Version, and testing through direct Lambda invocations and real Workflow runs. Production deployments should use immutable, qualified Lambda version ARNs to avoid replay and determinism risks, and organizations can adopt serverless gradually, run it alongside dedicated Workers, or use it as overflow capacity on the same Task Queue.

Temporal Workflows support resilient, long-running distributed applications but require deterministic code execution, making safe workflow updates essential as business logic evolves. The post contrasts code-level patching with Worker Versioning, which manages multiple Worker builds at the deployment layer so pinned workflows remain on their original version while auto-upgrade workflows can move to the current version. Temporal’s open-source Kubernetes Worker Controller automates this process through custom resources, creating versioned deployments, progressively ramping traffic, supporting autoscaling, and removing old Workers once pinned workflows have drained. The demonstrated CI/CD setup uses GitHub Actions and a self-hosted runner to build Docker images tagged with commit SHAs, apply WorkerDeployment resources to a local Minikube cluster, and trigger deployments automatically when relevant code changes are pushed. In the example, a greeting workflow already in progress continues using its original build after a code change, while newly started workflows use the new version, and the old deployment is removed after completion. Rollbacks can be initiated by redeploying a known-good image tag, with the controller recognizing it as an immediate rollback rather than repeating a gradual rollout.

Following its front-of-shirt sponsorship of Crystal Palace, Temporal is extending its support to Crystal Palace Women, framing the move as part of a broader commitment to equal access and opportunity. The announcement highlights rapid growth in women’s football in England alongside persistent disparities in funding, broadcasting, and career pathways, and compares these challenges with gender representation gaps in UK technology, where women comprise 20% of software engineers and 21% of senior tech leaders. Support for the women’s team will include investment in performance, sports science, recruitment, data, and audience development, while the company also plans to connect the team with its technology community through content and matchday experiences. The initiative will additionally emphasize opportunities for girls through the Palace for Life foundation and is presented as a long-term effort to help women in both football and technology gain greater resources and visibility.

Greg Haskins of Manetu describes how the same boundary between an AI agent and external systems can address both production durability and security concerns, including non-deterministic behavior, long-running human interactions, crashes, scaling, identity, prompt injection, credential exposure, and unauthorized tool use. Manetu’s AgentVisor uses Temporal’s Workflow model to make LangGraph-based agents durable without changing their application code, mapping each Agent Protocol Thread to a Temporal Workflow, Runs to Workflow Updates and Activities, and checkpoints to persisted Workflow state. This design allows agent processes and their latest checkpoints to be restored on new workers after failures, while stable Workflow IDs support location-independent access to long-lived conversations. To control state growth, the system retains only the newest LangGraph checkpoint and relies on Temporal replay for historical restoration, while handling human-input resumes differently from crash recoveries and carefully coordinating Continue-As-New rollovers. AgentVisor also uses the runtime boundary for host-issued identities, credential brokering that keeps real secrets outside agent environments, and transparent encryption of Temporal state through a Payload Codec, aiming to improve both operational resilience and agent trustworthiness.

Temporal can improve the reliability and efficiency of large, flaky test suites by treating independently executable tests as durable units of work rather than rerunning an entire CI job after a transient failure. In a reference implementation using Playwright, each test specification and browser combination is represented as a Temporal Activity, while a Workflow coordinates concurrent execution, preserves completed results, and retries only failed activities according to configured policies. Temporal does not fix broken tests, race conditions, or dependencies between tests, but it records progress across infrastructure failures, worker crashes, network issues, and intermittent browser problems. The approach keeps the existing test framework unchanged: Temporal code remains a wrapper that discovers test work, orchestrates it, and invokes normal commands such as Playwright, Go, Java, or other language-specific test runners. Safe concurrency requires tests to avoid shared resources or to isolate items such as ports, databases, directories, containers, and accounts. Activities can also be distributed among multiple Workers through a Temporal Task Queue, enabling scalable execution across CI runners or production infrastructure while allowing unfinished work to be reassigned if a Worker disappears.

During Temporal Cloud’s migration of usage and billing data storage to ClickHouse, coding agents proved most valuable for removing temporary scaffolding and outdated configuration rather than creating new code, largely because the team had documented cleanup needs and local rules in advance. Temporary migration code included distinctive comments explaining its purpose and required removal, allowing agents to locate and delete related packages, mocks, manifests, and environment settings efficiently. The team also placed durable, package-specific instructions alongside relevant code and linked them to regression tests, such as a rule requiring numeric values to be widened before aggregation to prevent overflow. However, written agent instructions can become obsolete when systems change, and one surviving file still refers to deleted components and invalid verification steps. The authors caution that incomplete prompt records and a lack of comparison data prevent claims that agents made the overall migration faster, while emphasizing that domain-specific knowledge remains essential.

LlamaIndex describes adopting Temporal, an open-source durable workflow execution platform, to replace increasingly complex RabbitMQ-based orchestration for LlamaParse, its document-parsing service that handles more than 130 file types and processes documents with varied requirements such as text extraction, OCR, vision models, and distributed CPU/GPU workloads. As document sizes and processing diversity grew, the previous system required custom mechanisms for state tracking, retries, heartbeats, fairness, backpressure, rate limits, and recovery from failed jobs, creating operational complexity and requiring overprovisioning. Temporal provides durable workflows and activities whose execution state persists across failures, enabling work to resume reliably and allowing workflows to coordinate through unique IDs, signals, leases, and resource-locking patterns. LlamaIndex built workflow-based semaphores and dedicated concurrency-limiter workers to manage permits, project-level job limits, and shared resources without relying on separate caches, databases, or limiter services. The company reports that the transition consolidated several supporting layers and enabled LlamaParse to scale to tens of millions of pages processed daily for its Batch API while allowing engineering teams to focus more directly on document-processing logic.