Diagrid’s Dapr 1.19 simplifies the workflows of AI agents and eliminates the need for a control plane service

Agent infrastructure startup Diagrid Inc. today released Dapr 1.19 with changes that reduce the overhead of persistent workflows behind long-running artificial intelligence agents.

Open source runtime operators can also delete one of their control plane services. Until 1.19, Dapr’s workflow engine provided the entire execution history to the worker at every step.

Everyone contributes to this story. With a job that runs for hours or splits into dozens of parallel tasks, the payload keeps growing, and Diagrid said this led to an increase in latency and memory usage. Now a worker keeps what it already contains and only receives new events.

This means that the data sent per step is independent of how long a workflow has been running. Workers still play the entire story locally. The smaller transfers only go to workers who declare their support, which in 1.19 means those based on Dapr’s Java and Python software development kits.

Mark Fussell, co-founder and chief product officer of Diagrid, said an agent “that runs all afternoon and splits itself into 50 tasks cannot transport its entire history through the network at every step.”

A second workflow change, an opt-in preview feature called WorkflowsFastPath, prevents every workflow event from being routed through a persistent schedule reminder. Wake-up calls and local activities will be conducted on site. The memory path remains as a backstop for recovery, and in Diagrid’s engine benchmarks, the mean step latency dropped from 553 milliseconds to 153 milliseconds with the feature enabled.

Previously, two services formed Dapr’s control plane, and a new mode in 1.19 removes Placement, which distributes the actor placement table. The remaining Scheduler service manages orders, reminders and workflow scheduling, already tracks cluster membership and pushes state to sidecars in the same way as Placement.

Yaron Schneider, co-founder of Dapr and also co-founder and chief technology officer of Diagrid, said the pair “grew into two services that do overlapping work.” When the mode is enabled, the scheduler takes over actor placement and the standalone placement service is no longer provided. This saves operators from having to have a stateful service and the Raft consensus cluster behind it. Reminders also become faster because the scheduler knows which sidecar belongs to each actor.

Scaling events are also less disruptive. A joining or leaving host paused placement across the namespace while Dapr rebroadcast its placement table. In the new mode, only the affected actor types will be blocked and updated, and the rest will continue to serve traffic.

New alpha building blocks in 1.19 are aimed at developers of AI applications. A Vector building block provides a portable application programming interface for storing and querying embeddings through similarity search, with Meilisearch being the first supported component.

The Binary Store streams large files such as model outputs in chunks, so they never have to remain entirely in memory. The initial components include Amazon S3 and Microsoft Azure, Google Cloud and Oracle object storage.

Developers working in .NET get Actors.Next, a new Actor SDK that generates proxies and registration code at compile time, reducing runtime overhead and allowing Actor apps to compile with Native AOT.

Dapr 1.19 is now available. WorkflowsFastPath ships disabled and scheduler placement mode is enabled via a Helm setting once everyone is running Sidecar 1.19, without requiring a sidecar restart to turn it on or off. Teams running Dapr on Kubernetes can manage the upgrade through Diagrid’s Dapr Ops dashboard, and the company said the tool typically supports new versions within 24 hours.

Image: Diagrid