Durable execution for agents

Why this conversation

An agent that works for a long time will, sooner or later, be running on a machine that dies. What happens to its half-finished work is the question Diagrid is built around. Mark Fussell is the company’s Chief Product Officer and co-founder, and the host recorded with him on 23 October 2026, the second day of AGNTCon + MCPCon North America in San Jose. He explains durable execution, why agents make it more urgent, how Diagrid proves what an agent did, and where it runs. The other conversations from that day are in the day-two article.

More than a few retries

Diagrid’s focus is what Mark calls durable execution. His starting point is a stateful business process that runs for a long time, such as an order form that sends out inventory and then moves through a series of steps, much like a state machine. If the process or the machine underneath it dies, durable execution recovers every state variable and everything that was executing, and the work continues where it stopped.

The host wondered whether retries alone would do. Mark separates two kinds of failure. In the first, an external system is not there, and the engine retries the call. The second is harder: a process that was running has itself failed. That needs snapshots and checkpoints of the running processes, and a way to detect that the service is down. External retries and process recovery are both required.

Why agents raise the stakes

This matters more now, he says, because the agents built today with LangGraph, Strands, Google ADK or Microsoft Agent Framework are long-running systems working in a loop. A RAG application that reads many documents is one example: if a single document cannot be read because of a network timeout, the work pauses and the business logic may retry it, but the whole machine can fail as well.

Recovery is also what stops an agent from repeating itself. If it has already written to a database or sent an email, restarting it from scratch would do those things a second time and leave a mess, and it would pay again for token calls it had already made. His way of putting the guarantee: run a thousand agents, attack and kill them with chaos testing, and every one of them recovers its state and carries on to completion.

Business agents, signed step by step

The host recalled leaving a task running in Claude Code and coming back to find it had stopped. Mark draws a line there. Coding agents have their own ways of coping with failure; Diagrid concentrates on business agents, the ones that run a company’s processes. One, built for a warehousing company, reads all of its incoming emails. Another was with Zeiss, the lens manufacturer, where an agent took the information from prescriptions, created the glasses and sent them out. “They need to be accurate as well.”

The other focus is security. Each agent gets an identity, which decides which agents it may talk to and which MCP servers it may use. The same identity lets Diagrid sign every step an agent takes. Logs and audits are still needed, and are very good for debugging, but they can be tampered with. When someone asks whether an agent really acted on their behalf and sent a given email, “we show that cryptographically that that’s actually what happened.”

Offline, at the edge, and how to start

A lot of Diagrid’s customers are in finance, where they build trading systems or plain business workflows; a lot of others are manufacturers. “One thing that we’re very proud of is that our solution will actually work completely disconnected offline.” It can run inside a manufacturing plant or in a company’s own cloud infrastructure. Asked about the edge, Mark says Diagrid is Kubernetes-first: a set of machines running a local Kubernetes cluster is enough. At the conference he had just spoken to a chip designer that wanted to optimise its manufacturing with language models and described exactly that situation.

The product is called Diagrid Catalyst. To try it, sign up for a free account at diagrid.io, where there are demos, scenarios and a guide to getting started. “We think it’s a fun name, you know, it catalyzes your company.” There is also Diagrid University, with courses hosted in a VM that run in the browser. His advice is to take the free courses, build a few applications, and check how well they survive failure.

When it comes out

The episode is edited and in the release order. Like every episode it goes out on YouTube first, then on Apple Podcasts, Spotify and Amazon Music.