Diagrid this week at the KubeCon + CloudNativeCon Europe conference announced that the 1.0 release of a framework for running artificial intelligence (AI) agents that is based on the open source Distributed Application Runtime (Dapr) project is now available.
Written in Python, Dapr Agents 1.0 makes it possible to deploy AI agents in a way that is easier to recover in the event of an outage or interruptions, says Diagrid CEO Mark Russell. That’s because the AI agent takes advantage of Dapr to make it possible to run it within a workflow rather than just in memory, he adds.
If there is a crash or disruption, execution steps such as the user input, intermediate steps, tool calls, and decisions are all persisted. The Dapr Agents framework runs every agent invocation as its own durable workflow, including interactions between agents. Agents can discover and orchestrate other agents dynamically at runtime by also incorporating the Secure Production Identity Framework for Everyone (SPIFFE) framework. “These capabilities can be federated across an entire agentic AI system,” says Russell.
A process simply resumes from exactly where it left off, not from the beginning or some arbitrary checkpoint set by a developer. Instead, any disruption is automatically detected to ensure processes are recovered in a way that ultimately burns fewer tokens, he notes.
In contrast, other frameworks for running AI agents in a production environment don’t provide the same level of resiliency, says Russell.
Other benefits include Model Context Protocol (MCP) integration, which enables dynamic discovery and invocation of external tools. Agents can also be invoked as child workflows in a fixed orchestration pattern. This is important when teams need predictable execution, auditability, and tight control over how specialized agents coordinate.
Agents automatically register themselves in a runtime Agent Registry for discovery by other agents and IT operations teams and are also automatically instrumented using OpenTelemetry tracing.
Originally started by Roberto Rodriguez from NVIDIA, Dapr Agents has grown rapidly since its initial release last year. The project has shipped over 20 releases that iterate on a core set of application programming interface (APIs) and durability primitives layered on top of Dapr.
Builders of AI agents can opt to use Dapr Agents to build AI agents or deploy Dapr underneath another AI agent framework to develop similar capabilities, notes Russell. The Dapr Conversation API also makes it possible to swap providers of large language models (LLMs) without code changes, with native clients for OpenAI, Azure OpenAI, NVIDIA, Hugging Face, and ElevenLabs all available.
Currently being advanced under the auspices of the Cloud Native Computing Foundation (CNCF), Dapr was originally developed at Microsoft to run as a container sidecar on a Kubernetes cluster. It provides a set of APIs for running distributed computing tasks such as service-to-service invocation, state management, and publish/subscribe messaging. That eliminates the need for each developer to write that same code over again for each application.
Agent framework competition is shifting from reasoning capability to operational durability, notes Mitch Ashley, vice president and practice lead for software lifecycle engineering at the Futurum Group. Dapr Agents 1.0 builds on a CNCF graduated runtime to position durable execution, automatic recovery, and SPIFFE-based agent identity as the production readiness baseline other frameworks will need to match, he adds.
For platform teams evaluating agent infrastructure, the core question is whether the execution substrate persists state, recovers from failure, and enforces identity across agent boundaries, says Ashley. “Teams deploying agents without these foundations will face the same reliability gaps that plagued early microservices adoption,” he says.
It’s still early days so far as the building and deploying of AI agents is concerned, but one thing is clear: how they are being created today will not be the same a few months from now as more hard-won experience is gained.
