Diagrid Catalyst 2.0 Brings Failure Recovery and Cryptographic Proof to AI Agent Workflows

Diagrid has launched Catalyst 2.0, a platform designed to make AI agent workflows more resilient by adding checkpoint-based recovery and cryptographic verification of execution history. The service integrates with ten major agent frameworks, including LangGraph, Microsoft Agent Framework, Google ADK, and others, allowing developers to embed durability into existing applications.

The Problem It Solves#

Multi-step agent runs are vulnerable to a common failure mode: when a long sequence of model and tool calls breaks partway through, the retry must re-execute every prior call unless checkpoints are in place. Catalyst addresses this by tracking each model and tool invocation as a durable workflow activity. When a run fails, execution can resume from the last successful step without repeating completed work, reducing wasted compute and latency.

Core Features#

Durable Execution Developers add a Diagrid package to their agent code. The platform then represents model and tool interactions as recoverable workflow activities, enabling intelligent retry logic at fine granularity rather than at broader graph boundaries. This approach works across cloud, on-premise, and air-gapped deployments, with agent building supported through a Python SDK and workflow SDKs available for .NET, Go, Java, JavaScript, Python, and Rust.

Verifiable History The verification layer draws from Dapr 1.18, which implements cryptographic signing of workflow history. Dapr hashes batches of events, chains the digests with signatures using the sidecar's SPIFFE-based identity, and signs the result. When state is loaded, the chain is validated; tampering—whether deletion, reordering, or modification—becomes detectable. External parties can verify histories against the Dapr Sentry trust anchor, creating an auditable record of what an agent actually did.

Early Adoption and Positioning#

ZEISS Group is among the first to use Catalyst, describing it as a "stable foundation" in a shifting landscape of AI models and frameworks. Yaron Schneider, Diagrid's co-founder and CTO, emphasized that organizations need "proof of what happened" when agents delegate work or call tools.

Catalyst enters a market with existing alternatives. LangGraph persistence checkpoints at graph boundaries, while Temporal and Restate offer replay-based durability for long-running workflows. Catalyst's differentiation lies in a unified Dapr-backed model that spans multiple frameworks and operates at the granularity of individual calls, rather than broader architectural layers.