Dapr Agents v1.0, released at KubeCon Europe on March 23, 2026, is the Cloud Native Computing Foundation's production-grade framework for running AI agents reliably in Kubernetes — bringing durable workflows, persistent memory, SPIFFE identity, and scale-to-zero to agentic AI workloads.
What Problem It Solves
Most AI agent frameworks are optimised for prototyping. When you run agents in production, you hit a different set of problems:
- Failure recovery: Long-running agents that fail mid-task have no way to resume — they start over, burning tokens and time
- State management: Agent memory is in-process; a pod restart loses everything
- Security: Agent-to-agent communication lacks workload identity and encryption
- Scaling: Idle agents consume resources even when doing nothing
Dapr Agents addresses all four by building on Dapr's distributed application runtime — the same battle-tested infrastructure used by thousands of Kubernetes workloads.
Why It's in Assess
Dapr Agents v1.0 was released just days ago and represents the first production-grade cloud-native agent framework from a major open-source foundation. It is compelling for:
- Platform engineering teams building internal agent infrastructure on Kubernetes
- Enterprise organisations that require workload identity, audit trails, and failure recovery
- Teams already using Dapr — Dapr Agents adds AI agent primitives with no new infrastructure
It's in Assess rather than Trial because the ecosystem is very new and most teams will be evaluating rather than deploying it to production immediately.
Key Features
Durable Execution
DurableAgent runs as a full Dapr Workflow instance — every invocation is persisted. If the pod crashes at step 7 of a 10-step task, the agent resumes from step 7, not step 1. This is the fundamental shift that makes long-running agents viable in production.
Persistent Memory Agent state survives pod restarts, deployments, and node failures. Works with 30+ state store backends — Redis, PostgreSQL, DynamoDB, Cassandra, and more.
Virtual Actor Model with Scale-to-Zero Idle agents are unloaded from memory but retain all state. When called again, they activate in ~3ms (tp90). This makes running hundreds of specialised agents economically feasible — you pay only for active computation.
SPIFFE-Based Workload Identity Built-in identity for agent-to-agent communication using SPIFFE standards — the same approach used for service mesh security in cloud-native systems.
30+ State Store Integrations Plugs into whatever infrastructure you already run — no new data stores required.
Origins & Governance
- Created by: Roberto Rodriguez at NVIDIA
- Co-developed with: NVIDIA, the Dapr open-source community, and enterprise end users
- Announced at: KubeCon + CloudNativeCon Europe, Amsterdam, March 23, 2026
- Governed by: CNCF (under the Linux Foundation umbrella)
- Real user: ZEISS Vision Care presented a production implementation at KubeCon — using Dapr Agents to extract optical parameters from unstructured documents at scale
Compared to Other Agent Frameworks
| Dapr Agents | OpenAI Agents SDK | LangGraph | |
|---|---|---|---|
| Durable execution | Yes (built-in) | No | No (external) |
| Scale-to-zero | Yes (virtual actors) | No | No |
| Workload identity | Yes (SPIFFE) | No | No |
| Best for | Production Kubernetes | OpenAI ecosystem | Complex graph logic |
| Licence | Apache 2.0 | MIT | MIT |
Getting Started
# Install Dapr CLI
wget -q https://raw.githubusercontent.com/dapr/cli/master/install/install.sh | /bin/bash
# Initialize Dapr locally
dapr init
# Install Dapr Agents
pip install dapr-agents
Key Characteristics
| Property | Value |
|---|---|
| Version | 1.0 GA |
| Governed by | CNCF (Linux Foundation) |
| Language | Python |
| Infrastructure | Kubernetes + Dapr |
| Licence | Apache 2.0 |
| Activation latency | ~3ms (tp90) with virtual actors |
| Release date | March 23, 2026 |