Restate raises $20M for durable execution. Why your AI agents need to survive a crash mid-task

What happens when your agent crashes on step 7 of 12? If you do not know, this one is for you.

What happened

Berlin based Restate raised a $20 million Series A led by Singular, with Redpoint Ventures and Capital One Ventures participating. Restate builds a durable execution engine, which records a program's progress so that if something fails mid-process, it can recover and carry on rather than start over or leave work half done.

Co-founder Stephan Ewen, a co-creator of Apache Flink, told TechCrunch the product "never was built for agents in the beginning" but turned out to fit the problems agents surface. In Restate's own announcement he writes that a single agent can make hundreds of model calls, tool calls, waits, retries and human approvals. Customers include Replit and Fortune 500 firms in financial services, per TechCrunch. The company is positioning against Temporal, which TechCrunch says raised $550 million at a $12.55 billion valuation earlier this month. Restate's post also argues durable actions are too expensive today, citing managed pricing often around $25 to $50 per million actions versus roughly $1 per million database writes.

My take

You do not need Restate to take the lesson. Most business automations I review have no answer for a failure halfway through. The CRM got updated, the email did not send, and the retry sends the invoice twice.

Ask these questions of any agent or long workflow:

1. If it fails at step 7, does a rerun repeat steps 1 to 6? Can each of those steps safely run twice?
2. Is progress saved somewhere, or only in memory?
3. When it waits on a human approval for two days, does it survive a restart?
4. Can you see exactly which step failed and resume from there?

n8n, Make and Zapier each give you some of this through execution logs, retries and error paths. For heavier agent work, a durable engine like Restate or Temporal starts to earn its place. Either way, idempotent steps (safe to run twice) are the cheapest insurance you can buy.

When I build systems for operations teams, the failure path gets designed before the happy path ships. If you want that thinking applied to your workflows, start at romielwillautomate.dev.