The ground beneath the agent
At Amazon Web Services in 2009, Maxim Fateev kept seeing engineers arrive with the same broken solution. Each team building systems that needed to execute every step exactly once — charging cards, reserving inventory, dispatching orders — had reinvented the same duct-tape framework: state flags in a database, dead-letter queues, retry logic bolted on after the fact, and hope. Fateev’s answer was Amazon Simple Workflow Service — the first attempt at AWS scale to give developers reliable coordination for long-running, asynchronous work.
SWF was powerful and difficult. Fateev and his colleague Samar Abbas eventually went separate ways: Abbas to Microsoft, where he built the Azure Durable Task Framework; Fateev to Google. Then in 2015 Uber opened a development center in Seattle and pulled them back together. The problem they had been circling for six years was waiting. Every engineering team at Uber was solving it independently — payments, dispatch, onboarding, each with its own state machine and retry infrastructure and edge cases. “In three years, we grew from zero to a hundred use cases within Uber,” Fateev would recall. Their internal engine was called Cadence. DoorDash, Coinbase, and HashiCorp started using it without being asked.
In 2018 they left Uber. In October 2019 they founded Temporal Technologies and released Cadence’s open-source successor, free from four years of accumulated migrations. The design was simple: record every step as an immutable event in an append-only log, then replay that log to reconstruct state after any failure. A workflow could run for minutes or months; if the process died mid-flight, Temporal would resume it exactly where it left off, without re-executing steps that had already succeeded. No double charges. No partial states. No flags.
Fateev had a pitch for skeptical engineers. “I’m not going to use a database. I’m a smart guy. I’ll just write and read files.” Nobody laughed — because every engineer had tried exactly that, and every one of them had eventually capitulated to a database. The analogy was a claim about inevitability: what the database ended for data persistence, durable execution would eventually end for execution.
The claim had to wait for the right pressure. That pressure arrived with AI agents — systems that coordinate dozens of tool calls over hours and cannot afford to lose state when a container restarts. On September 14, 2026, Temporal announced a $550 million Series E at a $12.55 billion valuation — more than double the $5 billion it had been worth in February. OpenAI’s usage had grown sixty-fold in under a year. The platform processed 1.9 trillion actions in August alone. “Durable execution,” said Venkat Venkataramani, OpenAI’s VP of Infrastructure, “is more than ever a core requirement for modern AI systems.”
An agent making fifty tool calls over two hours, if it dies on call forty-nine, has wasted the two hours — and possibly corrupted whatever real-world state it touched on the way. Samar Abbas, Temporal’s CEO, put it plainly: “Reliability has never been optional. AI has quickly raised the cost of skipping it.”
The architecture Fateev first sketched at Amazon in 2009 was not ahead of its time. It was waiting for software to get complicated enough to need it.
Sources
- Maxim Fateev on the OSS Startups Podcast — Temporal Blog — Fateev on building SWF at AWS, Cadence’s growth at Uber (“zero to a hundred use cases”), the database analogy, and founding Temporal.
- Temporal Technologies: Founding Story — Contrary Research — Founding timeline, founders’ careers at AWS and Microsoft Azure, key early customers, and funding history.
- Temporal Raises $550M Series E at $12.55B Valuation — Yahoo Finance — September 14, 2026 funding round: investors, OpenAI usage metrics, customer counts, quotes from Abbas and Venkataramani.