
CoreWeave launches Forge with agent execution and management tools
The AMW Read
Forge meaningfully expands an established GPU provider into agent execution and management, broadening infrastructure competition without demonstrating adoption or resolving the orchestration-moat debate.
CoreWeave launches Forge with agent execution and management tools
CoreWeave introduced Forge at Fully Connected 2026 in San Francisco, bringing development and management tools onto its GPU infrastructure. The platform combines Weights & Biases, OpenPipe, and Marimo with services spanning execution, monitoring, and evaluation. Agent Lens traces production agents' decisions and tool calls; generally available Sandboxes provide isolated CPU and GPU environments; and Registry tracks models, agents, and datasets with asset lineage. Serverless Sandboxes use Kata VM isolation and millisecond billing, while the on-cluster version runs within customers' existing CoreWeave Kubernetes Service at no incremental cost.
The launch expands CoreWeave's role from supplying compute to managing the workflows that consume it. In the AI infrastructure market, execution environments, observability, and asset management give providers additional ways to embed themselves in customers' operations. Forge also reaches into agent orchestration, where detecting failures and controlling tool execution matter beyond model capability. Its cloud-agnostic positioning and portable registry formats could ease migration, but moving integrated workflows may still involve operational work. That creates a potential retention advantage; the launch alone does not establish customer dependence or production reliability.
Builders should evaluate Forge against the cost and complexity of assembling these services separately. CoreWeave reports better failure detection and lower issue-resolution costs for Agent Lens, but those comparisons need validation on buyers' own workloads. A concrete procurement test is whether tracing, sandbox isolation, and lineage tracking simplify incident diagnosis while preserving a workable path to another cloud.