Summary of "OpenAI + @Temporalio : Building Durable, Production Ready Agents - Cornelia Davis, Temporal"

Summary of “OpenAI + @Temporalio : Building Durable, Production Ready Agents - Cornelia Davis, Temporal”

Key Technological Concepts and Product Features

1. OpenAI Agents SDK Overview

Enables building agentic AI applications where Large Language Models (LLMs) have agency to decide application flow.
Available in Python and TypeScript.
Core concept: an agentic loop where the LLM continuously decides actions, including invoking tools.
Supports tool integration, guardrails, and handoffs between agents.
Simplifies creating agents by abstracting complex flow control and tool invocation.
Includes tracing and debugging features.

2. Temporal Overview

Open-source distributed systems platform focused on durability and reliability.
Acts as a backing service providing durable state management, retries, event sourcing, and orchestration.
Widely used in production by companies like Snapchat, Airbnb, Pizza Hut, Taco Bell, and OpenAI (CodeX and image generation).
Supports multiple programming languages (7+ officially, including a recent Swift SDK).
Key abstractions:
- Activities: Units of work (e.g., external API calls) that may fail and require retries.
- Workflows: Orchestrations of activities representing business logic.
Provides automatic retry policies, state persistence, and fault tolerance.
Enables long-running workflows (minutes to years) with efficient resource management.
Workers are multi-threaded processes that execute workflows and activities, pulling tasks from event queues.
Durability ensures no loss of progress on crashes or network failures, avoiding redundant LLM calls (saving token costs).

3. Integration of OpenAI Agents SDK with Temporal

Temporal and OpenAI collaborated to integrate the Agents SDK with Temporal workflows and activities.
Integration brings durability and scalability to otherwise ephemeral agentic loops.
Activities are wrapped as tools using a provided activity_as_tool function that generates JSON tool descriptions automatically (no internal or private APIs needed).
The integrated system supports:
- Durable LLM calls with retry policies.
- Dynamic activities that can be invoked by name at runtime.
- Agentic loops running inside Temporal workflows for fault tolerance.
- Visibility and tracing through Temporal UI and Agents SDK tracing features.
Required changes to the OpenAI Agents SDK, including making the runner class abstract to support Temporal’s durability model.
Supports complex orchestrations including handoffs between microagents and parallel execution.

4. Agentic Loop Demo Highlights

Demonstrated a weather alert agent that queries weather APIs via tools implemented as Temporal activities.
Showed how the LLM decides when to invoke tools and how tool outputs feed back into the conversation.
Demonstrated durability by killing the worker process mid-execution and showing the workflow resumed exactly where it left off without redoing LLM calls.
Showed Temporal UI displaying detailed workflow and activity execution, retries, and event sourcing.
Highlighted dynamic activity invocation, allowing flexible tool usage without tight coupling.

5. Orchestration and Microagents

Agents SDK supports orchestrating multiple small agents (“microagents”) each with their own agentic loops.
Two orchestration methods:
- Simple chaining by passing results between agents.
- Using handoffs where an agent changes context to another agent without spawning a new loop.
Temporal supports parallelism, signals, queries, updates, and long waits, enabling complex agent orchestrations.
The microagent pattern parallels microservices architecture benefits: modularity, scalability, and independent deployment.

6. Additional Features and Roadmap

Current limitations:
- Native streaming support for agents is not yet available but is a top priority.
- Large payload storage (passing data by reference) is also being developed.
Customers have built streaming solutions on top of Temporal.
Support for additional LLM providers like Claude and Gemini is in progress.
Temporal is open-source with permissive licenses (mostly MIT).
Temporal Cloud offers a managed SaaS service with multi-region durability and persistence backed by relational DB or Cassandra.
Temporal does not currently host workloads (workers); users deploy and manage those themselves.
No native incident management integrations, but users commonly integrate with Kubernetes or custom alerting solutions.
Temporal is exploring voice agent use cases.

7. Developer Experience and Best Practices

Activities should be idempotent to handle retries safely.
Latency added by Temporal calls is minimal (tens of milliseconds), suitable for most agent applications.
Temporal allows developers to focus on business logic (“happy path”) without handling retries, failures, or state persistence.
Encourages modular code by separating tools from agent logic.
Ecosystem includes an AI cookbook with example patterns and workflows, plus a community code exchange.

Guides, Tutorials, and Resources Mentioned

Demo Repositories:
- Temporal AI Cookbook with agentic loop examples (branch: agentic-loop-de).
- Integration code between OpenAI Agents SDK and Temporal in the contrib directory of the Python SDK.
Workshops and Notebooks:
- Basic Jupyter notebook tutorial for OpenAI Agents SDK and Temporal (not used in this advanced session).
Documentation and Blogs:
- Temporal docs: docs.temporal.io with an AI cookbook section.
- Blog post describing the OpenAI Agents SDK and Temporal integration.
- Pydantic blog on durability integration with agent frameworks.
Community and Events:
- Temporal conference “Replay” in May at Moscone with discounted registration.
- YouTube channel with webinars on human-in-the-loop agents, MCP servers, and advanced Temporal use cases.
Hiring Notices:
- Temporal is hiring for engineering and developer advocate roles focused on AI applications.

Main Speakers and Sources

Cornelia Davis – Developer Advocate at Temporal, primary presenter.
Johan – Head of AI Engineering at Temporal, contributed engineering leadership and answered technical questions.

In summary: This presentation detailed how Temporal’s distributed systems platform brings durability, scalability, and fault tolerance to OpenAI’s agentic AI applications via a deep integration with the OpenAI Agents SDK. It showcased practical demos, explained core abstractions (activities, workflows), and highlighted the power of combining agentic loops with Temporal’s event sourcing and orchestration capabilities to build production-ready, durable AI agents.