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

Operators of the open-source runtime can also drop one of its control-plane services. Until 1.19, Dapr’s workflow engine shipped the entire execution history to the worker on every step.

Each one adds to that history. On a job that runs for hours or fans out into dozens of parallel tasks, the payload keeps growing, and Diagrid said that pushed up latency and memory use. Now a worker keeps what it already holds and gets only new events.

That makes the data sent per step independent of how long a workflow has been running. Workers still replay the full history locally. The smaller transfers go only to workers that advertise support for them, which in 1.19 means those built on Dapr’s Java and Python software development kits.

Mark Fussell, Diagrid’s co-founder and chief product officer, said an agent “that runs all afternoon and splits into 50 tasks can’t haul its entire history across the network on every step.”

A second workflow change, an opt-in preview feature called WorkflowsFastPath, stops pushing every workflow event through a durable scheduler reminder. Wake-ups and local activities are handled in place. The reminder path stays as a backstop for recovery, and in Diagrid’s engine benchmarks median step latency fell from 553 milliseconds to 153 milliseconds with the feature on.

Two services have made up Dapr’s control plane until now, and a new mode in 1.19 retires Placement, which distributes the actor placement table. The surviving Scheduler service handles jobs, reminders and workflow scheduling, and it already tracks cluster membership and pushes state to sidecars the same way Placement does.

Yaron Schneider, a co-creator of Dapr who is also Diagrid’s co-founder and chief technology officer, said the pair “matured as two services doing overlapping work.” With the mode on, the Scheduler takes over actor placement and the standalone Placement service is no longer deployed. That spares operators a stateful service and the Raft consensus cluster behind it. Reminders get faster too, because the Scheduler knows which sidecar owns each actor.

Scaling events are also less disruptive. A host joining or leaving used to pause placement across the entire namespace while Dapr re-sent its placement table. In the new mode, only the affected actor types are locked and updated, and the rest keep serving traffic.

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

Binary Store streams large files such as model outputs in chunks so they never have to sit in memory whole. Its first components cover Amazon S3 and the object stores from Microsoft Azure, Google Cloud and Oracle.

Developers working in .NET get Actors.Next, a new actor SDK that generates proxies and registration code at compile time, which cuts runtime overhead and lets actor apps compile with Native AOT.

Dapr 1.19 is available now. WorkflowsFastPath ships turned off, and the Scheduler placement mode is enabled through a Helm setting once every sidecar runs 1.19, with no sidecar restart needed to switch 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 releases within 24 hours.

Image: Diagrid