TechPrufer
Get started
Temporal
Temporal logoTemporal · Feature trail

ML Workflow Orchestration

18 posts · Mar 2024 – Aug 2026

What this trail covers

Temporal is being used to orchestrate complex AI/ML processes, including end-to-end ML pipelines and data engineering flows. This includes managing state machines, handling retries and rollbacks, and coordinating tasks across different infrastructure, such as GPU servers. The platform's durable execution model ensures reliability and allows for faster iteration by developers. Julep AI leverages Temporal to build multi-step AI workflows, enabling AI agents to remember past interactions and execut. This post details how Temporal and Chainlit can be integrated to build a Durable RAG system that provides reliable, efficient, and scalable conversational AI, addressing challenges in workflow orchestration, scalability, reliability, and user interaction for RAG models.

All Temporal trails
  1. Everything in its place: Making agents write correct Temporal applicationsThis post details how to leverage AI agents and Temporal's developer resources (Skills, Samples, MCP) to build robust Temporal applications. It emphasizes providing agents with the right context (Temporal Developer Skill, SDK Samples Repositories, Temporal Knowledge Base MCP server), guardrails, and clear objectives for successful development. The core idea is to improve the quality and correctness of AI-generated Temporal code by feeding the agents curated and canonical information sources.
  2. Durable RAG and agents: MongoDB and Temporal doing it better togetherThis post details a reference architecture for building durable RAG systems and AI agents by combining MongoDB Atlas for data storage (including embeddings and vector search) with Temporal for reliable workflow orchestration. It explains how Temporal's Durable Execution ensures data ingestion pipelines and agent reasoning loops survive failures, retries, and restarts without losing state. The post outlines a concrete use case of a RAG knowledge base over Temporal documentation, detailing the data ingestion process (chunking, embedding, indexing) and the agent's interaction with the vector store. It highlights Temporal's role in providing at-least-once guarantees for data ingestion, eliminating the need for Kafka, and enabling idempotent updates through workflow IDs. The agent side demonstrates how Temporal orchestrates the agent's reasoning loop, tool calling, and security controls, ensuring the agent resumes from its last state upon failure.
  3. Chaining Activities — from text to vectorsThis post details a specific ML workflow orchestration pattern: chaining Temporal Activities to process text data. It demonstrates how to embed text using an on-premises model, cache vectors in Redis, store them in a vector database (Pinecone), update a database flag (Postgres) to mark records as embedded, and perform a sanity query. Key technical contributions include the explicit definition of the workflow orchestration with discrete activities, the use of Redis as an external storage layer for embeddings (Claim Check pattern), the implementation of an 'embedded flag' pattern for reconciliation between Postgres and Pinecone, and a lightweight smoke test for the vector retrieval path. It also highlights the importance of idempotency at each layer (Postgres case_id, Pinecone upsert, embedded flag) and the safe retry of activities.
  4. How XY builds an AI agent orchestration platform for healthcare with TemporalThis post details how XY built an AI agent orchestration platform for healthcare using Temporal. It introduces a YAML-based Domain Specific Language (DSL) executed by a single generic `DSLWorkflow` class, enabling AI-powered workflow generation via a Planner Agent that translates natural language into executable YAML. The platform utilizes an Activity Factory for runtime assembly of customer-specific activities and a deferred Continue-As-New mechanism for infinite batch processing. It highlights the use of dedicated task queues and specialized workers for precise infrastructure scaling and emphasizes the benefits of Temporal's vocabulary for aligning engineering and product teams.
  5. Inside the Google ADK and Temporal integrationThis post details the integration of Temporal with Google's Agent Development Kit (ADK). The integration allows ADK agents to leverage Temporal's durable execution by running each LLM call and tool call as a separate Temporal Activity. This provides granular retries, improved visibility in the Temporal UI, and more robust failure handling compared to running the entire agent within a single Activity. Key technical contributions include the `TemporalModel` wrapper for rerouting LLM calls to Activities, the `activity_tool` helper for defining tools as Activities, and modifications to ADK's time and UUID functions to ensure determinism within Temporal Workflows. The post also discusses the handling of MCP tools and the serialization of context objects.
  7. Building observable AI agents: Temporal now integrates with BraintrustThis post details the integration of Braintrust, an AI observability platform, with Temporal. It explains how to instrument Temporal Workflows and Activities to trace LLM calls, manage prompt versions via Braintrust's `load_prompt` function, and leverage Braintrust's tracing capabilities for debugging and cost analysis of AI agents. The integration involves wrapping the OpenAI client with Braintrust's `wrap_openai` and adding the `BraintrustPlugin` to Temporal Workers and Clients. The example demonstrates a deep research agent workflow that benefits from both Temporal's durable execution and Braintrust's LLM observability.
  8. From prototype to production-ready agentic AI solution: A use case from Grid DynamicsThis post details Grid Dynamics' migration of a deep research agent from a LangGraph/Redis/Kafka architecture to Temporal. Key technical contributions include: demonstrating how Temporal's durable execution simplifies state management by making state an integral part of the workflow rather than a separate, fragile object; showcasing the elimination of thousands of lines of custom retry and error handling code by leveraging Temporal's declarative RetryPolicy configurations for Activities; and highlighting how Temporal's event history automatically and durably persists state updates, reducing the need for manual state management and debugging complex issues related to caching and stale state.
  9. Durable multi-agentic AI architecture with TemporalThis post details how Temporal can be used to build durable multi-agentic AI systems. It introduces two key patterns: Agent Routing and Task Delegation. A concrete example is provided for a detection and repair system for order management, demonstrating how to use Temporal Workflows and Activities to orchestrate AI agents for analysis, plan creation, and plan execution. The post also covers integrating conversational agents via the Model Context Protocol (MCP) and outlines the common steps for building agentic applications.
  10. Building production-ready generative AI: How Temporal supercharges Google's Gemini and VeoThis post details how Temporal's Durable Execution and Activity management are used to build a production-ready video generation system leveraging Google's Gemini and Veo models, addressing challenges like stateful long-running processes, transient failures, costly re-runs, and visibility/debugging. It showcases an open-source example using Gemini for scene generation and Veo for video clip generation, highlighting the use of structured output from LLMs and Temporal's automatic retries for resilience.
  11. Durable Execution meets AI: Why Temporal is ideal for AI agents & Generative AI AppsThis post details how Temporal's Durable Execution and Activity model are ideal for building resilient AI applications and agents, by providing robust orchestration for LLM invocations, tool execution, and state management, effectively addressing the distributed systems nature of AI applications.
  12. Building an agentic system that’s actually production-readyThis post provides a practical guide to building production-ready agentic AI systems, building upon the concept that agentic systems are distributed systems in disguise. It defines key terms like AI, Goal, Tool, User, Agent, and Agentic System, and introduces advanced concepts like Multi-Agent systems and Agents as Tools. The post illustrates the agentic flow (Interaction, Decision, Action) orchestrated by Temporal Workflows, and details the challenges in building these systems for production, such as unreliability of components, non-determinism of AI, complexity, and security. It strongly advocates for Temporal's Durable Execution as the solution for building stateful, fault-tolerant, and observable agentic systems, and provides a simplified code example of an agentic workflow framework using Temporal Signals, Activities, and workflow logic for planning and executing tools, including user interaction for confirmation and input gathering.
  13. Making friends with agents: A mental model for Agentic AI applicationsThis post introduces a mental model for building agentic AI applications, focusing on the event loop, LLM-tool interaction, and durable execution of these agents. It details the components of an agent (prompt, tools, execution mechanism, LLM input update) and emphasizes the need for a durable event loop and durable invocation of LLMs and tools. The post also discusses the language interfaces for LLMs and tools, and how to prepare for tool invocation by gathering and structuring inputs, potentially involving user interaction and LLM-driven validation. Finally, it outlines the components of LLM input for subsequent turns, including the goal, tools, example conversation, context instructions, and conversation history, highlighting the role of prompt engineering in managing these elements for effective agentic application development.
  14. What are Agentic AI Workflows? Scalable & Durable WorkflowsThis post details how Temporal can be used to build agentic AI workflows, addressing limitations of current frameworks in durability, scalability, and integration. It highlights Temporal's strengths in developer velocity, observability, scheduled execution, durable and resilient execution (handling LLM probabilistic nature and real-world failures), long-running and stateful capabilities, human-in-the-loop support, flexible integration with any LLM/database/service, and centralized orchestration. It provides a typical workflow example and emphasizes Temporal's role as a robust foundation for evolving AI agents.
  15. Efficient GPU resource management with TemporalThis post details how a technology company is using Temporal to build an extensible platform for managing long-running GPU resource management workflows. The platform was built in three months and leverages child workflows. The team is migrating parts of the platform to use Temporal Nexus for improved extensibility and integration. They also developed an automation pool for smaller, ad-hoc tasks. The post highlights the benefits of Temporal Cloud over self-hosting for managing infrastructure overhead and enabling developers to focus on application logic.
  16. Julep’s vision for the future of AI WorkflowsThis post details how Julep AI uses Temporal to build and manage complex, multi-step AI workflows. It highlights Temporal's role in orchestrating tens of thousands of tasks across different systems, ensuring reliable execution and retries. Specific use cases include video editing automation and personalized content creation, showcasing Temporal's ability to handle long-running, complex processes. The post emphasizes Temporal's expressiveness, its live UI for observability, and its deterministic workflow design with Python integration as key enablers for Julep's AI agent platform.
  17. 9 Real-World Generative AI Use Cases Powered by TemporalThis post details nine real-world generative AI use cases powered by Temporal, including video content processing and translation, conversational AI and call center transcription, synthetic data generation and processing, AI platforms for predictive analytics and customer insights, market intelligence and cryptocurrency data enrichment, AI-powered HR workflows, AI infrastructure and resource optimization, state tracking in generative AI workflows, and research and applied AI workflow orchestration. It emphasizes Temporal's role in coordinating complex, long-running, stateful workflows for these applications, highlighting benefits such as reliable performance, cost savings, and faster development.
  18. Temporal Boosts Video Processing by 600% at The Washington PostThis post details how The Washington Post's Arc XP division used Temporal to revamp their media workflows, specifically for an automated thumbnail selection system. By leveraging Temporal's durable execution and multi-language support, they achieved a 600% runtime improvement, reducing processing time from 80 seconds per minute of video to about 10 seconds for a 4-5 minute video. This was accomplished by distributing workloads across AWS instances. The Replay feature was highlighted as a key benefit for support staff in debugging and incident management.
  19. Durable RAG with Temporal and ChainlitThis post details how Temporal and Chainlit can be integrated to build a Durable RAG system. It explains how Temporal's durable execution, scalability, and resource optimization features address the challenges of workflow orchestration, scalability, and reliability in RAG models. It also highlights Chainlit's role in enhancing user interaction with pre-built components for chatbots, data visualizations, and feedback systems, enabling rapid prototyping and deployment of conversational agents. The integration aims to create a system that is both powerful and scalable, handling complex, long-running workflows and user interactions effectively.
  20. Amazon Bedrock with Temporal: rock solidThis post details the integration of Temporal Workflows with Amazon Bedrock to build a production-grade AI chatbot. It demonstrates how Temporal's durable execution handles stateless LLMs by storing conversation history within workflows, manages long-running conversations by allowing workflows to wait indefinitely without consuming compute resources, and provides event history for debugging. The post provides code examples for a basic workflow and activity using the Temporal Python SDK and AWS Bedrock, showcasing how to execute activities, manage retries with default exponential backoff, and run a worker. It highlights the benefits of guaranteed execution, state management, and simplified debugging for Gen AI applications.
  21. ML Workflows with Temporal: Optimizing AI and data engineeringThis post details how Temporal can be used to orchestrate model training, testing, and deployment, and manage complex data pipelines. It highlights the use of Workflows and Activities for reliable execution, automatic retries, and state management. The post also discusses optimizing GPU costs by running intensive workloads only when needed through Task Queues, and orchestrating multiple AI pipelines together.
  22. 9 ways to use Temporal in your AI WorkflowsThis post details nine ways Temporal can be applied to AI workflows: workflow orchestration for AI pipelines, scalable and reliable ML model training, distributed data processing, continuous learning and model deployment, experimentation and versioning, efficient use of GPUs, scaling AI operations, event-driven and asynchronous execution, and observability and debugging. It also provides links to Temporal SDKs and sample projects for getting started.