Restate Raises $20 Million to Help AI Agents Recover From Interrupted Work

Recovering from a crash is part of the infrastructure Restate sells—not a repair each application developer must build. On September 30, 2026, the startup announced a $20 million Series A led by Singular, with Redpoint Ventures and Capital One Ventures participating. The funding backs its push to make that protection routine across backend software, including the many steps inside AI agents.

Keeping progress when a process stops

Restate provides an execution engine designed to keep multi-step workflows resilient to crashes and network interruptions. Its approach is called durable execution: the runtime—the software managing a program’s execution—records progress, retries work and recovers partially completed operations after a failure.

The company’s funding announcement describes agents that can make hundreds of model and tool calls, wait for callbacks, retry operations and seek human approval. Those sequences bring the same coordination problems found in conventional workflows into the agent’s repeated cycle of deciding and acting.

Restate’s argument is about where recovery belongs. Rather than protect only the workflow surrounding an agent, it wants individual model calls, tool calls and guardrail checks recorded within the agent loop. The company says this produces a consistent execution history for sensitive, long-running or expensive tasks.

State can also outlive a single workflow. Restate’s Virtual Objects store information under an identifier, such as a chat-session ID, and allow one write handler at a time for each identifier. That gives developers a way to manage persistent agent context or sessions alongside durable functions.

A recovery engine built around a log

Restate built its own replicated log, consensus and failover handling around RocksDB for local storage and object stores for durable snapshots. The log keeps a persistent record of execution; RocksDB follows it to maintain the information needed to coordinate ongoing work. The company credits that design with fewer moving parts and better efficiency.

It also reverses a common workflow pattern. Instead of putting tasks in queues for workers to fetch, Restate pushes requests to deployed functions. Its streaming connections and suspension protocol are designed to support both quick steps and code that waits for a day. That requires explicit flow control, rather than relying on workers to slow their intake.

The funding thesis rests on removing three barriers to using durable execution more often. Restate identifies them as:

• Cost: The company puts common managed-service pricing at $25–$50 per million durable actions, versus roughly $1 per million database writes or queue events. It acknowledges that a durable action does more work.
• Latency: Restate argues that reducing the delay per durable action from more than 25 milliseconds to under five makes adding another protected step easier to justify.
• Deployment: Its clustered, dependency-free binary is intended to make self-hosting practical for sensitive applications that must keep data inside a network boundary.

Customer demand—and the hiring it funds

Restate presents Replit as an example of durability moving inside the agent process, rather than requiring a separate workflow orchestrator between activities. The higher action count measures architectural use, not a tenfold improvement in agent performance.

Its customer examples extend beyond agents. Restate says DOSS replaced a BullMQ-based system for enterprise workflows and subworkflows, improving failure handling, latency and debugging. An unnamed Fortune 500 bank also adopted it for financial workflows requiring consistency across regions.

Co-founder Ewen told TechCrunch that Restate had closed multiple six- and seven-figure customer contracts over the preceding few months. The company plans to use the new capital to hire engineers and go-to-market staff, and enlarge its San Francisco Bay Area office to be closer to current and potential customers.