Temporal Launch & Learn: San Francisco
This week, I attended the Temporal Launch & Learn conference in San Francisco. The conference focused on how developers can use Temporal to build reliable, long-running GenAI applications.
Conference agenda
Here are some of the sessions I attended:
Cornelia Davis presenting the Foundations of GenAI Applications & Temporal workshop
This session introduced the foundations of building GenAI applications using Temporal, blending LLM-based workflows with distributed systems reliability. It emphasized how Temporal simplifies GenAI orchestration by handling retries, durability, and execution flow behind the scenes.
The morning covered:
Fixed-flow: calls LLMs at defined steps, processes outputs deterministically.
Agentic-loop: gives the LLM autonomy to determine subsequent actions.
Both have practical value; not every system needs full autonomy.
Calls made via APIs (e.g., GPT-4/5, Claude).
Importance of context windows: models are stateless, so context must be preserved across calls.
Early examples used the Completions API, later transitioning to the Responses API for tool-use and planning.
Jupyter notebooks were used to run Python examples inside GitHub Codespaces (no local setup required).
The first exercise walked through prompting an LLM (give me five facts about elephants) and outputting the result as a PDF report.
Participants practiced calling LLM APIs, chaining results, and using a small generate_pdf() utility to simulate downstream processing.
Follow-up exercises demonstrated creating .env files, adding OpenAI keys, and working through additional notebooks covering foundations and adding durability to workflows.
AI systems are "distributed systems on steroids."
Temporal's core abstractions (workflows, activities, signals, and queries) form the backbone of durable GenAI orchestration.
Temporal provides out-of-the-box recovery for failed API calls, retries for rate limits, and consistency across human or system interruptions.
"Don't feel like you must give agency to the LLM; fixed flows still have tremendous value."
"These AI systems are distributed systems on steroids."
Ryan Albert (Speakeasy) presenting how Temporal is used to power MCP servers at Speakeasy
Ryan Albert's session showcased how Speakeasy uses Temporal to power MCP (Model Context Protocol) servers, a standardized way for LLMs to access tools and APIs.
MCP (Model Context Protocol) defines how models access external tools, similar to OpenAPI but built for LLM integrations. It standardizes tool schemas, parameters, and interactions.
Temporal powers asynchronous and durable tasks behind MCP servers.
Enables complex, long-lived tool invocations without losing state.
Used for workflows like job execution, event-driven processing, and approvals.
Built a live TypeScript-based MCP server wrapping the Temporal Client API.
Exposed tools for querying active Temporal workflows and fetching workflow metadata.
Showed live deployment and integration with Claude via VS Code.
Dynamic MCP servers to reduce context window size using semantic tool discovery.
Human approval workflows: using Temporal signals for action verification before running critical tools.
Task-based long-running workflows coming to the MCP spec (e.g., getTasks, listTasks, cancelTasks RPCs).
MCP and Temporal together make long-running, fault-tolerant LLM agents practical.
"Temporal's logical workflows map perfectly to MCP's new long-running task spec; it's a match made in heaven."
"Now the robots are reading our docstrings." (on auto-generating MCP tool schemas from TypeScript docstrings)
"With Temporal, you can stop worrying about process boundaries; workflows just keep going."