Diagrid's Catalyst enables AI agents to resume from failures without starting over

AI agents often shine in controlled demonstrations but can stumble when deployed to handle real-world tasks. Diagrid's latest release of Catalyst 2.0 addresses two critical pain points: recovery from failures and compliance-ready audit trails for high-stakes applications.

Durable execution across multiple frameworks#

Starting Tuesday, Catalyst provides a resilience layer that sits beneath popular agentic frameworks including LangGraph, Microsoft Agent Framework, Google's Agent Development Kit, and OpenAI Agents SDK. Rather than asking developers to switch to a new framework entirely, Diagrid's approach runs underneath existing tools and converts each model call, tool invocation, and handoff into steps within a durable workflow.

When an agent crashes mid-execution, Catalyst allows it to pick up from exactly where it left off. According to Diagrid co-founder and CTO Yaron Schneider, this matters enormously in production: "If the agent gets a prompt and it chooses to run 100 tools for the job and it fails at the 99th, it really needs to start back up from 99."

The underpinning uses Dapr (Distributed Application Runtime), an open-source project the Diagrid team helped create at Microsoft. For each supported framework, Diagrid provides a runner that intercepts the agent's execution loop in real time, registering operations as workflow activities. In LangGraph, for instance, developers pass a compiled graph to Diagrid's DaprWorkflowGraphRunner. Catalyst records inputs and outputs, and when the system recovers from a crash, it replays the orchestration using stored results instead of re-executing completed steps.

LangGraph already offers its own persistence layer at checkpoint boundaries, and Schneider acknowledges that capability. Diagrid's advantage is uniform support across 10+ frameworks, with granular recovery down to individual model and tool calls, eliminating the need for custom recovery logic per framework.

Signed and immutable execution history#

Catalyst 2.0 also introduces workflow-history signing, a feature Dapr introduced in version 1.18, now extended to all supported agent frameworks. This creates a tamper-evident record that appeals particularly to regulated industries.

The system maintains a signed ledger of every input, output, and external system interaction. Dapr computes SHA-256 digests across batches of workflow events, chains each digest to the previous one, and signs the result using the Dapr sidecar's SPIFFE identity. If anyone attempts to modify, remove, or reorder stored events later, the verification chain breaks, revealing the tampering.