Anthropic introduces Workflow and the multi-agent operations question gets a first answer
A coordination layer for agentic AI: declarative workflow definitions, agent-to-agent messaging, and a planner. The framing is workflow orchestration. The substance is closer to operating-system primitives.
On 24 February 2026, Anthropic launched Workflow, a coordination product for multi-agent AI systems. The product introduced declarative workflow definitions in YAML, agent-to-agent messaging primitives, and a planner that decomposed high-level user goals into agent-level subtasks. As Anthropic's launch post framed it, Workflow was to multi-agent AI what container orchestration had been to microservices: an opinionated runtime that took the operational complexity off individual application teams.
The launch followed Dispatch (February 2026) and Skills (October 2025) in what was clearly a coherent product strategy. Skills had given organisations a way to package capabilities. Dispatch had given them a way to run them durably. Workflow gave them a way to compose them across multiple agents. As MIT Technology Review observed in coverage that week, Anthropic was building the operational stack for agentic AI as a parallel to the cloud-software stack of the early 2010s.
What Workflow does, and does not, replace
Workflow did not replace traditional workflow-orchestration products like Airflow, Dagster or Temporal. It coordinated AI agents specifically, with primitives sized for agent-to-agent interaction. As The Information's analysis pointed out the following week, the runtime model assumed agents were durable but expensive (ten-second responses being normal), small in number per task (single-digit agents per workflow being typical), and in need of decision-checkpointing rather than orderly state-machine progression.
Container orchestration was the cloud-software primitive. Agent coordination is the AI-software primitive.
Whether Workflow becomes a durable Anthropic product or simply pulls competing labs into shipping similar products is one of the open commercial questions of mid-2026. The early read, as of the writing of this piece, is that the design pattern, rather than the product itself, is the more strategically important thing. The same way Kubernetes outgrew its origin at Google in 2015 and 2016, the multi-agent coordination primitive will likely outgrow Anthropic's specific product.
Discussion