Diagrid: durable execution for business agents

Mark Fussell is Chief Product Officer and co-founder of Diagrid, and the co-creator of Dapr, the Distributed Application Runtime, a Cloud Native Computing Foundation graduated project. Luca Berton talked to him at AGNTCon + MCPCon North America 2026 in San Jose on 23 October 2026 and asked what Diagrid does. The answer is two words: durable execution.

What durable execution means

“When you have a long-running, stateful business process, quite frequently it has to be able to survive failures in production.” His example is a classic business workflow: an order processing form, inventory sent out, a series of steps like a state machine. If the machine or the process dies, you want its state fully recovered and the work to carry on from where it left off.

Retries are part of it. A call to an external system that isn’t there has to be retried, and Diagrid’s engine does that. The harder part is when the process itself fails. That takes snapshotting and checkpointing of the running processes, and detecting that the service has failed. “So we focus not just on retries, but process recovery and the durability.”

Why agents need it

This becomes much more important with agents, he says. LangGraph, Strands, Google ADK and Microsoft Agent Framework all run long-running agentic systems in a loop. A RAG application that processes many documents runs for a long time; “the last thing you want to do is that agent fails and you need to recover its state.”

Recovering is also about not doing things twice. An agent may already have run database transactions or sent an email. “If you restart your agent again and it does those again, now you’ve got a whole complete mess you have to deal with.” Recovered state means no repeated actions and no token calls wasted on work already done. His promise: “Imagine you’ve got a thousand agents running and you do chaos testing on all of them … We will guarantee all those agents recover their state, carry on and run to completion.”

Business agents, not coding agents

Luca mentioned coming back to a task left running in Claude Code and finding it stopped. Mark draws a line there. With coding agents you can deal with failures in other ways; Diagrid focuses on the business agents that run business processes. One is a warehouse agent that reads all the incoming emails, a long-running agent that has to recover its state if the machine dies. Another was built with Zeiss, the lens maker: an agent that takes the information from glasses prescriptions and creates the glasses that are sent out. Such processes “have to be sustainable, they have to be recoverable.”

Identity, and proof of every step

The other focus is security. Diagrid gives agents an identity, which decides which agents an agent may talk to and which MCP servers it may call. With that identity, every step an agent takes is cryptographically signed. “So you can prove that an agent sent an email at this time, wrote a database call here.” Logs and audits are still needed, for debugging among other things, “but audits and logs, they can still be tampered with.” If someone asks whether an agent really acted on your behalf and sent that email, the signatures show it.

Disconnected, on Kubernetes

Diagrid works with many financial customers, building trading systems or business workflows, and with many manufacturers. One thing he is proud of: it runs completely disconnected, offline, in a manufacturing plant or in your own cloud infrastructure. It is “a Kubernetes-first native system”, so a set of machines running a local Kubernetes cluster at the edge is enough.

How to start

Diagrid is built on open source Dapr. The product is Diagrid Catalyst, at diagrid.io, with a free account, demos and a getting-started guide. There is also Diagrid University: free courses that run in a VM in the browser. His suggestion: take the courses, build some applications, “and test them for resiliency and durability.”

More

The episode with Mark Fussell is being produced and goes out on YouTube first. The day it was recorded: AGNTCon + MCPCon North America 2026: day two.