Diagrid announced Catalyst 2.0 on 28 July 2026, introducing failure recovery and cryptographic verification capabilities for agents constructed with frameworks including LangGraph, Microsoft Agent Framework, Google ADK and Dapr Agents. The release ships with ten framework integrations and addresses a persistent challenge: when long multi-step agent runs fail late in a sequence, retries without checkpointing force developers to re-pay for every model call already completed. Catalyst converts model and tool calls into durable workflow activities, enabling interrupted runs to pick up without repeating finished work.
The platform supports agents built with a Python SDK, while Catalyst's workflow SDK documentation lists support for .NET, Go, Java, JavaScript and Python, with the open-source Dapr 1.18 SDKs also including Rust. The ten framework integrations cover LangGraph Deep Agents, AWS Strands, OpenAI Agents SDK, Claude Managed Agents, CrewAI and Pydantic AI alongside the four named earlier. Developers add a Diagrid package to an existing agent application, and Catalyst runs in cloud, on-premise and air-gapped environments. The verification model originates from Dapr 1.18, which hashes batches of workflow-history events, links each digest to the preceding signature and signs the result using the Dapr sidecar's SPIFFE-based identity, making deleted, reordered or modified history detectable when workflow state is loaded.
Yaron Schneider, Diagrid co-founder and CTO and chair of the Agentic AI Foundation's Workflows and Process Integration Working Group, said organisations need "proof of what happened" when agents call tools or delegate work. The announcement cites optics manufacturer ZEISS Group as an early user, with Wendelin Niesl, its head of end-to-end core application engineering, saying Catalyst provides "the stable foundation we can rely on" in a rapidly evolving landscape of AI models and frameworks. Diagrid claims Catalyst can deliver up to ten times the performance of open-source Dapr and support millions of concurrent agent workflows, though the announcement doesn't specify whether that multiple refers to throughput, workflow starts per second or latency, nor which workload, hardware or Dapr configuration was measured.
Catalyst isn't the first durability option for these frameworks—LangGraph persistence records checkpoints at graph superstep boundaries, while Temporal and Restate provide replay or journal-based execution for long-running applications and can already host agent workflows. What distinguishes Catalyst is a single Dapr-based recovery and attestation model spanning several frameworks, applying durability to individual model and tool calls rather than graph boundaries. The architectural decision therefore hinges on where teams want recovery, identity, audit evidence and framework integration to reside. Teams should note that signing is disabled by default in Dapr 1.18, gated behind the WorkflowHistorySigning feature flag and dependent on mTLS, with daprd refusing to start if signing is enabled and mTLS is off—and it's a one-way decision per workflow, with no retroactive signing of existing history and in-flight unsigned workflows needing to complete or be purged before enabling it more widely. Cryptographic attestation proves the integrity and provenance of recorded history, not that an agent made a correct decision, a tool returned accurate data, or every external side effect was captured, and practitioners should also assess retry handling for non-idempotent tools, storage and latency overhead, certificate rotation, independent verification, and which capabilities belong to open-source Dapr versus commercial Catalyst. The commercial model spans a free cloud tier, dedicated-cloud and bring-your-own-cloud plans sized by concurrent workflows, and a custom-quoted Enterprise Server edition for on-premise and air-gapped deployments. Organizations betting on agentic workflows will need to weigh whether centralized durability infrastructure simplifies operations enough to justify vendor lock-in, or whether framework-native persistence keeps flexibility without adding another dependency layer.