Diagrid gives failed AI agents a way to resume
The New Stack Frederic Lardinois ● Covered by 2 sources
Diagrid's Catalyst 2.0 lets AI agents pick up where they crashed instead of restarting from scratch. It also creates a tamper-evident log of every step — handy once regulators start asking questions.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Every AI agent demo looks flawless until it meets real traffic, and then the thing that ran 98 tool calls without a hitch dies on call 99 and, without help, starts the whole job over. Diagrid's answer, launched Tuesday as Catalyst 2.0, isn't a new framework to learn. It's a layer that sits underneath the agent frameworks people already use — LangGraph, Microsoft Agent Framework, Google's Agent Development Kit, OpenAI Agents SDK, and others — and quietly turns every model call, tool call, and handoff into a step in a durable workflow.
The mechanics come from Dapr, the open source runtime Diagrid's own team helped build while at Microsoft, and its workflow engine. For each framework Diagrid supplies a runner that hooks into that framework's execution loop and registers its operations as workflow steps. Co-founder and CTO Yaron Schneider puts the payoff bluntly: if an agent chooses to run 100 tools for a job and fails at the 99th, it needs to resume at 99, not restart at zero. In a LangGraph app, for instance, a developer compiles the graph as usual, hands it to Diagrid's DaprWorkflowGraphRunner, and Catalyst records inputs and outputs so Dapr can replay the orchestration after a crash, feeding back stored results instead of rerunning completed work.
None of this is entirely new territory. LangGraph already has its own persistence layer that checkpoints at superstep boundaries and its Agent Server carries a durable task queue. Diagrid's pitch is less about beating that and more about consistency: the same recovery model across more than 10 frameworks, applied down to individual model and tool calls, so teams stop writing bespoke recovery logic for every framework they touch. Schneider says LangGraph is by far the most common framework among Diagrid's customers, with AWS Strands and Microsoft Agent Framework showing up too, and the rest filling out a long tail that's cheap enough to support anyway.
The second piece of Catalyst 2.0 may matter more to enterprise buyers than the resume feature does. Diagrid has brought the workflow-history signing introduced in Dapr 1.18 into every supported agent framework. Schneider describes it as keeping a diary — logging inputs, outputs, and which systems the agent talked to. Under the hood, Dapr computes a SHA-256 digest over batches of workflow-history events, chains each digest to the one before it, signs the result using the Dapr sidecar's SPIFFE identity, and checks that chain every time it loads workflow state. Tamper with an event — modify it, delete it, reorder it — and the verification breaks. Customers can bring their own certificates and keep the encrypted history readable even after they've stopped running Catalyst.
Diagrid is clearly aiming this second feature at regulated industries. CEO Mark Fussell says financial executives he's spoken with view the absence of a verifiable record as a real blocker to putting agents into sensitive workflows, and the EU's AI Act gives the pitch extra weight: Article 12 requires high-risk AI systems to support automatic event logging so operators can trace behavior and monitor deployments. Fussell frames Catalyst as something that runs alongside a cloud provider's existing identity, evaluation, and observability tooling rather than replacing it, and it can be deployed as a hosted service or inside a customer's own environment, air-gapped setups included. Diagrid hasn't said what any of this costs.
My take — AI-written commentary, not fact-checked reporting
The recovery trick is the obvious, boring, correct fix nobody wanted to build framework-by-framework, and that's exactly why it'll get adopted quietly without headlines. The signed-history piece is the more interesting bet — it's Diagrid reading the EU AI Act's logging requirement as a sales opportunity rather than a compliance headache, and that's smart, because most vendors are still treating Article 12 like paperwork instead of product. Leaving pricing out of the announcement is the one part that smells like a company that knows enterprise procurement teams will ask anyway.
Read more about this at: The New Stack
