Temporal
Durable execution platform that keeps long-running agent and backend workflows alive across crashes, restarts and outages, with SDKs in Python, Go, Java, TypeScript and .NET.
Overview
Temporal is durable execution, a dull name for a useful guarantee. You write ordinary functions in ordinary code. Temporal journals every step, input and result, and if the process holding that workflow dies, another worker picks it up and replays the journal to the exact point of failure. State survives crashes, deploys, restarts and multi-day waits. The server is MIT-licensed and the SDKs cover Python, Go, Java, TypeScript and .NET.
For agents this matters more than most framework arguments do. An agent that calls six tools, waits on a human approval and runs for an hour is a distributed systems problem wearing an LLM costume. Temporal had already solved that shape of problem for payments and order processing when the agent crowd arrived: OpenAI, Replit and Lovable run agent workloads on it, and the OpenAI Agents SDK integration reached GA on the Python SDK in March 2026. The same year brought Workflow Streams for durable token streaming and External Storage for payloads too large to sit inside event history.
The honest caveats are cost and discipline. Workflow code must be deterministic, so no wall-clock reads, no random values and no direct network calls inside the workflow function, and violations surface later as replay divergence instead of an immediate error. Cloud pricing meters Actions rather than executions, which rewards tight workflows and punishes polling loops. Self-hosting removes the invoice and adds a cluster to operate. Teams already running Kubernetes and Postgres at scale will find that trade acceptable. Smaller teams should start on Essentials at $100 a month and watch the Action counter before they commit.
Key Features
- ✓ Durable execution: every step, input and result is journaled, so a worker that dies mid-run is replaced and the workflow replays to the exact point of failure
- ✓ Open source SDKs for Python, Go, Java, TypeScript and .NET, with community SDKs covering PHP, Ruby and Rust
- ✓ OpenAI Agents SDK integration went GA on the Python SDK on 23 March 2026, with sandbox environment support in public preview
- ✓ Workflow Streams, added in 2026, gives durable token streaming and application-level updates built on the Signal and Update primitives
- ✓ Nexus for durable workflow-to-workflow and service-to-service calls, GA on Python and in preview on TypeScript and .NET
- ✓ External Storage offloads large payloads to S3 or a custom driver, which starts to matter once agent context blobs outgrow the event history
Where it holds
- • It fixes the failure mode that quietly kills agent demos in production: a 40 minute run that dies at minute 38 and has to start from zero.
- • Language spread is unusual. A Go backend team and a Python ML team can share one workflow engine without either side rewriting.
- • The people running agents at real volume already pay for it. OpenAI, Replit, Lovable, Block and Nordstrom are named customers, and the February 2026 Series D put $300M in at a $5B valuation.
Where it breaks
- • Action-based billing is the sharp edge. Workflows that poll or fan out generate far more Actions than a napkin estimate suggests, and you find out monthly.
- • Self-hosting is free and not cheap. You are operating a cluster, a persistence layer and a version upgrade path, and most teams underestimate all three.
- • It is not an agent framework and does not claim to be. You still pick LangGraph, Pydantic AI or the OpenAI Agents SDK for the reasoning loop, then wrap it.
- • Determinism rules inside workflow code catch people out: no clock reads, no random values and no direct network calls, and the punishment is a replay divergence rather than a clean error at write time.
My Take
February 2026 brought a $300M Series D at a $5 billion valuation, and the pitch attached to it was specific: agents that survive failure. That framing is right. Nothing here writes prompts or picks tools, it simply guarantees that a workflow which crashed at step nine of twelve wakes up at step nine with its variables intact, possibly on a different machine. The catch is the meter, since Temporal Cloud counts Actions rather than runs and Essentials includes only one million a month for its $100, so model your volume for a fortnight before you migrate anything real.
Quick Info
- Pricing:
- freemium
- Openness:
- Open source
- Licence:
- MIT
- Starting at:
- The server is MIT-licensed and free to self-host, which is a real option and a real ops burden. Temporal Cloud meters Actions: Essentials from $100 a month, Business from $500, Enterprise custom, with the first million Actions included, the next 5M at $50 per million and the rate stepping down to $25 per million in the 100M to 200M band. New accounts get $1,000 in credits and startups under $30M raised can apply for $6,000. Watch the counter rather than the sticker price: community reports put a single agent workflow anywhere from a handful of Actions to fifty, so naive per-million maths understates the bill. Verified on temporal.io/pricing, August 2026.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Temporal holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Temporal
Convex
freemiumReactive TypeScript backend where queries, mutations and durable workflows run inside the database. Agent and Workflow components handle long-running tool loops.
LangGraph
open-sourceLow-level orchestration framework for stateful, long-running LLM agents, with checkpointing, human-in-the-loop pauses and a managed deployment tier through LangSmith.