AI
Why Agent Frameworks Leave Reliability to You
Popular agent frameworks like LangChain and LangGraph handle reasoning, tool calling, and LLM abstractions. They do not handle distributed systems problems. Their answer to failure is checkpoints: JSON-serialize each step’s output, save it to a database. But checkpointing alone does nothing when the agent crashes. The framework hands failure detection and recovery back to the developer. That means writing a workflow engine, handling duplicate executions across instances, and building a recovery process that must itself survive failure. Most teams have not done this yet.
Dapr Agents 1.0: Durable by Default
Dapr Agents was contributed by Nvidia and released as stable 1.0 the day before this talk. The framework wraps standard agent patterns (name, role, instructions, LLM provider) with a durable workflow engine underneath. State saves to any of Dapr’s 30-plus supported databases. A kill-nine crash mid-execution leaves the workflow intact in the database. When the process restarts, it resumes at the exact step it stopped. No re-curl, no re-prompt. The framework also saves conversation history alongside workflow position, so no prior context needs recomputation.
The Kill-Nine Demo: Automatic Recovery in Practice
Schneider ran a local weather agent, sent it a “what is the weather in Amsterdam” query, then killed the process before it finished. No graceful shutdown, no cleanup. He restarted the agent without re-sending the curl request. The agent picked up exactly where it stopped and returned the answer. Dapr workflows also apply configurable retry policies: set a circuit breaker, retry up to five times before marking the workflow failed, then use the Dapr management API to inspect and restart failed instances in bulk.
Multi-Agent Coordination With Four Lines of Code
The multi-agent demo used two agents, Frodo and Sam, communicating over Dapr pub/sub backed by Redis. Swapping Redis for Kafka requires changing one YAML file, not the agent code. Each agent runs with its own Dapr sidecar for full process isolation. The orchestrating agent decides which agent handles each step and passes messages iteratively until the workflow completes. Up to 50 agents can be connected this way. Every message carries OpenTelemetry trace data, giving a full audit log in Zipkin or any compatible backend.
Dapr as Foundational AI Infrastructure
Beyond Dapr Agents, the broader Dapr building blocks apply directly to AI workloads. Pub/sub gives at-least-once delivery guarantees between agents running in separate pods. State management covers over 30 databases, not just the Redis, Postgres, and in-memory SQLite that most frameworks ship with. The Conversation API adds PII obfuscation without an external service, stripping credit card numbers and addresses before they reach the LLM and scrubbing any that come back out. Access policies let operators define which agents can reach which MCP servers, creating auditable trust boundaries at runtime.
Notable Quotes
No agent framework can do this today. None will do this because, again, they are not focused on distributed systems challenges. Yaron Schneider · ▶ 14:37
my position today is that none of those frameworks are actually production ready. Um not for enterprise scale. Yaron Schneider · ▶ 15:46
That thing is like a five-year-old’s brain on an ayahuasca trip. Yaron Schneider · ▶ 07:00
Dapr workflows really guarantees that only one instance of every workflow is going to be executing. Yaron Schneider · ▶ 18:42
Key Takeaways
- Dapr Agents 1.0 shipped stable the day before KubeCon EU 2026, with Nvidia among the first production users.
- Checkpoint-only frameworks require developers to build their own failure detection and recovery systems from scratch.
- Dapr workflows guarantee exactly-once execution across any number of concurrent agent workflows without extra infrastructure.
About the Speaker(s)
Yaron Schneider is CTO and co-founder of Diagrid. He co-created the CNCF projects Dapr and KEDA while at Microsoft and led engineering architecture for serverless container platforms at scale. He is a steering committee member of the CNCF and was recently elected chair of the workflow working group inside the CNCF AI Foundation.