Enable scalable structured AI agent coordination with Dapr Workflows | ODFP976

by Kendall Roden

· Breakout Session · 15 min · 3 min read

Applied AI AI Infrastructure

TL;DR

Kendall Roden demonstrated Dapr's stable workflow API (1.15) running on Catalyst, a managed platform that drops the Kubernetes sidecar requirement. An order processor workflow coordinates inventory, payment, and shipping from any compute target (serverless, container, VM) with a single HTTP trigger. Roden also showed Dapr Agents, a new multi-agent framework where an LLM orchestrator routes tasks across specialized agents over Dapr pub/sub, with max iterations capping planning rounds and automatic retries preserving state.

Dapr’s Platform Architecture and Diagrid’s Enterprise Layer

▶ Watch (00:38)

Dapr started at Microsoft in 2019 to simplify distributed systems and graduated as a CNCF project. It delivers portable building-block APIs (service invocation, pub/sub, state, durable workflows) through a language-agnostic sidecar. Diagrid, founded in 2021 by Dapr’s co-creators, fills the production gap with three products: D3E (a Dapr distribution with namespace isolation and reduced Kubernetes permissions), Conductor (automated upgrades, certificate rotation, and real-time security advisories), and Catalyst (a fully managed Dapr platform that moves the sidecar off your pod and exposes Dapr APIs over the network, reachable from containers, VMs, or serverless functions).

Durable Workflow Execution Without Infrastructure Overhead

▶ Watch (05:33)

The Dapr workflow API stabilized in version 1.15. On Catalyst, a Python app registers a workflow runtime, chains activity calls for inventory checks, payment, and shipping, then triggers a run with a single HTTP POST. The application can be a serverless function, a container, or a VM. It only needs the Dapr SDK and network access to Catalyst. The order processor workflow demo (05:55) shows the Catalyst dashboard generating the execution graph in real time, with inputs, outputs, and retry history visible with no extra tooling.

“This is real orchestration of a real business process.” — Kendall Roden

Multi-Agent AI Coordination with Dapr Agents

▶ Watch (09:57)

Dapr Agents is a new framework built on top of Dapr for production-grade multi-agent AI systems. Each agent wraps an LLM with tool access, memory, and messaging. A separate LLM orchestrator reads the agent registry from Catalyst’s key-value store, routes tasks to the right specialist, and limits planning rounds via a max iterations parameter. The Lord of the Rings multi-agent demo (10:38) shows this live: character agents collaborate over Dapr pub/sub, with Catalyst tracking every tool-calling workflow and LLM round. If any agent fails, Dapr retries without losing workflow state.

“It’s also remarkably simple, with just a few lines of code” — Kendall Roden

Notable Quotes

This isn’t a simulation or mocked up flow. Kendall Roden · ▶ 9:15

This is real orchestration of a real business process. Kendall Roden · ▶ 9:17

It’s also remarkably simple, with just a few lines of code Kendall Roden · ▶ 14:48

Key Takeaways

Dapr’s workflow API hit stable in 1.15; Catalyst removes Kubernetes sidecar requirements entirely.
The Dapr Agents framework separates LLM orchestrator logic from specialized character agents, enabling retries without lost state.
Agent code reduces to importing one class and configuring role, goal, and messaging store parameters.

Filed under

Applied AI AI Infrastructure