Temporal for AI agents: what you get and what you still build
Ayo Adeniran · Oct 30, 2026 · 6 min read
AI agents rarely fail in interesting ways. A worker restarts. A provider times out halfway through a tool loop. A manager answers the approval on Thursday instead of Tuesday. Each of these loses the agent's place, and with it the work.
Temporal's answer is durable execution, and it is a good one. On October 30, 2026 Temporal made its AI framework integrations generally available, so agents written with the OpenAI Agents SDK, Google ADK, LangGraph and others can run on it. If you are weighing Temporal for agents, here is what you get, what it costs, and what remains yours to build.
How Temporal works
Temporal's promise is simple to state: "Once a Workflow starts, it runs to completion, whether that takes a second or a year. If the process running it crashes, the Temporal Service hands the work to another process, which rebuilds the state of the execution and resumes at the point where it stopped."
It does that with an Event History, "a complete, ordered record of everything that happened in a Workflow Execution." A new worker fetches the history, replays the workflow code against it, and carries on. Two rules follow:
- Workflow code must be deterministic, "because Temporal reruns it to rebuild state after a failure."
- Anything that talks to the outside world belongs in an Activity: calling an API, writing to a database, sending an email, charging a card.
Running workflows receive Signals and Updates and answer Queries. You write them in one of eight SDKs: .NET, Go, Java, PHP, Python, Ruby, Rust or TypeScript. The Temporal server is open source under the MIT licence, or you can use Temporal Cloud.
Temporal for AI agents
The new integrations let teams "run agents built with different frameworks in Temporal Workflows, with the same durable execution model across them." The generally available set covers Google ADK, the OpenAI Agents SDK, AWS Strands, LangGraph, LangSmith, the Vercel AI SDK and the Gemini SDK.
With the OpenAI Agents SDK, for example, the agent loop, tool choice and handoffs run inside the workflow, and model calls run as Activities, so they retry durably and are not repeated during replay. A helper, activity_as_tool, wraps ordinary Activities so the agent can use them as tools.
What it costs
| Temporal Cloud | Price |
|---|---|
| Pay as you go | From $50 per million Actions, no base fee |
| Volume, self-service | Down to $25 per million Actions |
| Active storage | $0.042 per GB-hour |
| Retained storage | $0.00105 per GB-hour |
| Business support add-on | From $500 a month, including 2.5 million Actions |
Actions are the billable operations between your application and Temporal Cloud, including workflow, activity, timer, signal, query and schedule operations. An agent that makes many model and tool calls makes many Activities, so estimate on your real loop. Self-hosting the open-source server swaps this bill for the cost of running it.
What stays in your code
Temporal is general-purpose by design. It makes your code durable, and its own rule, that anything touching the outside world belongs in an Activity, tells you where the agent-specific work goes:
- Model calls are Activities, written by you or supplied by a framework integration. Which model serves each step, and what to do when a provider is down rather than slow, is logic inside them.
- Tool calls are Activities too, along with the credentials they use.
- A person's answer arrives as a Signal or an Update. Showing them the question, and sending their answer back, happens in your application.
- Workflow code stays deterministic, so prompts, branching and agent logic are written to replay cleanly.
- Workers are yours to run, whether you self-host the server or use Temporal Cloud.
For a platform team, that is the appeal: nothing is imposed. For a team that wants a working approval process this quarter, it is a fair amount of building.
Temporal and Momor Workflows side by side
| Temporal | Momor Workflows | |
|---|---|---|
| You write | Deterministic code in one of eight SDKs | YAML, or a C# builder; a canvas in the Designer |
| Durable record | Event History, replayed through your code | An append-only record of every step; a resumed run continues from the first unfinished step |
| Long waits | Workflows can run for a year; answers arrive as Signals | A Wait step holds the run's place, holding no worker, until a person answers |
| Model calls | Activities you write, or via framework integrations | Every call through Momor Gateway, with provider failover |
| People | Your application, sending Signals or Updates | Momor's Chat, the Workflows app, the API, or an MCP client's form |
| Tools | Activities you write | MCP servers' tools as steps; approval before a gated tool |
| Hosting | Self-host the MIT server, or Temporal Cloud | Momor |
| Pricing | Per million Actions, plus storage | Under an agreement |
Where Momor fits
Momor Workflows takes the durable-execution idea and builds the agent parts in.
The record comes first. Before the engine asks a model or calls a tool, it records that it is about to, and it records the answer when it arrives. After a restart, a run continues from the first step that did not finish; a finished step never runs again. Each turn runs the version of the workflow that was current when it opened, start to finish.
People are part of the definition. A Wait step asks a question and the run keeps its place until someone answers, and the time spent waiting is not counted against the run's working time. A tool from a connected MCP server can be marked as needing approval, so the call waits until the run's owner allows or refuses it.
Models without plumbing. Every model call goes through Momor Gateway. A step can name the models it may use, and a call moves to another provider when one fails before answering.
Tools and assistants over MCP. Connected MCP servers' tools become steps, signed in through OAuth as the organization or as each member. And your chosen workflows are offered as MCP tools to claude.ai, ChatGPT and other clients, where the person signs in and each run belongs to them.
Work that notices change. What a run relied on is recorded in Momor's knowledge graph. When one of those facts changes, the runs that relied on it are asked to look again.
What to weigh: Temporal runs any code in eight languages, and you can host it yourself; Momor runs workflows built from its step kinds, on Momor. Momor has no schedule, email or webhook trigger yet, while Temporal Actions include schedule operations. Choose Temporal when you want a durable foundation for anything. Choose Momor when the thing you need is an AI process with people in it, working this quarter.
Read the Workflows quickstart, see how runs and recovery work, or talk to us.