Learn how to create resilient AI research agents using Python, OpenAI, and Temporal. This tutorial covers implementing durability, handling failures seamlessly, and maintaining workflow progress without re-running expensive calls.

🔗Resources & Links
Repo with solution: https://github.com/temporalio/edu-deep-research-tutorial-template/tree/video_tutorial

👉Text Tutorial: https://learn.temporal.io/tutorials/ai/deep-research/?utm_campaign=awareness-nikolay-advolodkin&utm_medium=video&utm_source=youtube

🤝Connect with Nikolay: https://www.linkedin.com/in/nikolayadvolodkin/  

🕐 Timestamps
 00:00 - Introduction to the problem of process disruptions in AI agents
 00:48 - Overview of the OpenAI deep research agent setup
 01:14 - Observability and workflow visibility with Temporal
 01:43 - Cloning the repo and environment setup instructions
 02:12 - Configuring API keys and running the research agent locally
 03:09 - Demonstration of research workflow and disruption scenarios
 03:39 - Common failure points and solution overview with Temporal
 04:08 - Deep dive into agent pipeline components and their functions
 04:34 - Visual map of research pipeline: triage, planning, search, and report
 05:03 - How OpenAI SDK simplifies agent orchestration
 05:44 - Chaining agents: plan, search, and write in sequence
 06:13 - Durability through Temporal's integration with OpenAI SDK
 06:43 - Introducing Temporal for state persistence and error recovery
 07:42 - Building the interactive research manager class
 08:06 - Methods for planning, searching, and report generation
 09:04 - Managing clarifications and user input with workflows
 09:33 - Implementing triage and clarification questions
 10:00 - Automating research pipeline execution based on user interaction
 10:30 - Limitations of in-memory state management, solution with Temporal workflow
 10:55 - Persisting session state and indefinite human waiting with Temporal
 11:24 - Defining data classes and workflow signals for communication
 11:52 - Instantiating manager inside the workflow for durability
 12:22 - How durable activities are executed within Temporal context
 12:50 - No manual retry or checkpointing needed with Temporal
 13:19 - Pattern for human-in-the-loop interactions: start, clarify, resume
 13:46 - Temporal's handling of paused workflows without resource usage
 14:15 - Workflow handlers for get status, provide clarification, start research
 14:45 - Pausing workflow until user input, no polling or timeouts required
 15:14 - Difference between queries, signals, and updates in Temporal workflows
 15:43 - Initiating research session and storing clarification questions
 16:10 - Handling user inputs and updating workflow state with signals
 17:05 - Main run method: waiting for user input and progress management
 17:33 - Workflow resumption after user input, intelligent waiting
 18:00 - Cleaning up old in-memory research manager
 18:27 - Creating the worker to run Temporal workflows
 18:57 - Configuring worker with OpenAI plugin and retry policies
 19:27 - Setting timeout and retry parameters for durability and fault tolerance
 19:55 - Connecting to Temporal client with OpenAI plugin support
 20:25 - Running the worker and managing server restarts seamlessly
 21:21 - How Workflow handles session IDs, persistence, and resuming work
 21:51 - Full demo of the agent executing with failures and restarts
 22:21 - Explaining the role of Temporal client for starting and managing workflows
 22:50 - Frontend API: starting research, querying status, submitting answers, fetching reports
 23:18 - Handling workflow handles and querying current state without disruption
 24:16 - Endpoint implementations: get status, submit clarification answers, fetch results
 25:33 - Pattern for session management with session IDs and workflow handles
 26:03 - Running server, worker, and UI setup: a step-by-step guide
 26:30 - Demonstration of workflow execution, clarification, and report generation
 27:29 - Troubleshooting and fixes for demo success
 28:28 - How to handle server restarts during research sessions
 29:27 - Workflow continues seamlessly after worker restart with Temporal
 30:26 - Showcasing retry attempts, handling failures, and retry policies
 31:26 - Final report review and insights on the demo success

Temporal's platform is a central hub for developers interested in distributed systems, workflow orchestration, and fault-tolerant computing. Through articles, tutorials, and documentation, Temporal offers insights into building resilient and scalable distributed applications. Developers can learn about workflow automation, state management, and handling failure scenarios using Temporal's workflow engine and coordination APIs.

Temporal

A step-by-step tutorial on building a multi-agent deep research system using Python, the OpenAI Agents SDK, and Temporal. The tutorial starts with a basic multi-agent pipeline (triage, clarification, planner, search, and writer agents) and then adds Temporal to make it durable. Key problems solved include crash recovery without re-running expensive LLM calls, human-in-the-loop waiting without burning resources, and automatic retries with configurable backoff. The Temporal integration works via the OpenAI agents plugin, which intercepts runner.run calls and executes them as durable activities. The tutorial covers workflow definitions, query/update/signal handlers, worker configuration, and a FastAPI server connecting the UI to Temporal workflows.

Build a Deep Research Agent | Python, OpenAI, Temporal