Diagrid Catalyst 2.0 enables AI agents to resume from failure across multiple frameworks

Catalyst 2.0 Enhances AI Agent Resilience

Diagrid has released Catalyst 2.0, introducing a durable execution and attestation layer for AI agents. This update aims to improve the resilience of agents, particularly in high-stakes applications, by allowing them to recover from failures without losing progress.

Cross-Framework Compatibility

Catalyst 2.0 integrates with multiple popular AI agent frameworks, including LangGraph, Microsoft Agent Framework, Google's Agent Development Kit, and OpenAI Agents SDK. Diagrid states that Catalyst operates beneath these existing frameworks, converting agent model calls, tool calls, and handoffs into steps within a durable workflow.

Resumption from Interruption

The core functionality of Catalyst 2.0 allows an AI agent to resume execution from its last completed step if it is interrupted. This eliminates the need for the agent to restart the entire process from the beginning, saving computational resources and time, especially in complex, multi-step tasks. For example, if an agent fails at the 99th out of 100 tool calls, it can restart from that 99th step.

Underlying Technology

Catalyst is built upon the open-source Distributed Application Runtime (Dapr) and its integrated workflow engine. Diagrid provides a runner for each supported agent framework, which intercepts the framework's execution loop and registers operations as workflow activities. This mechanism records the inputs and outputs of model and tool calls, enabling Dapr's workflow runtime to replay orchestrations after a crash by returning stored results for completed activities.

Distinction from Existing Solutions

While some frameworks, like LangGraph, offer their own persistence layers for resuming from successful steps, Diagrid's Catalyst provides a consistent execution model across more than ten frameworks. It extends durability to individual model and tool calls without requiring developers to adopt a new agent framework.