CoreWeave Forge: The Enterprise AI Loop, Unified
CoreWeave announced Forge on September 30, 2026 at Fully Connected, its annual AI cloud conference in San Francisco. The launch marks a decisive pivot: the GPU cloud provider that built its business on raw compute is now competing directly with MLOps platforms, agent development tools, and AI infrastructure vendors on a single connected platform. Forge brings Weights and Biases Models, OpenPipe post-training, marimo notebooks, and four new CoreWeave services under one roof, and it arrives at a moment when the gap between deploying an agent and improving one has become the most expensive problem in enterprise AI.
What Happened, and Why It Matters
For the last two years, enterprises have been piecing together agent infrastructure from disconnected components: a model provider here, a tracing tool there, a fine-tuning service somewhere else. Each handoff between tools creates latency, lost context, and manual reconciliation work. The engineering cost of running that loop, rather than the cost of model inference, now dominates total AI spend at many large deployments.
CoreWeave’s answer is to collapse the loop. Forge is designed so that a production trace from Agent Lens feeds directly into a training run in Weights and Biases Models, which feeds into a Serverless RL job, which feeds back into a new deployment, all without leaving a single account or reconciling data between systems.
“Engineering teams need to understand those gaps, identify the signals that matter, improve the next version and measure whether the change worked under real operating conditions,” said Chen Goldberg, EVP of Product and Engineering at CoreWeave. “None of this has to live in a disconnected system.”
What Ships With Forge
CoreWeave Forge ships with ten components across the AI loop:
CoreWeave Agent Lens (New)
Agent Lens is the most operationally significant new service. It provides end-to-end agent tracing at production scale, analyzing tens of millions of traces automatically. Rather than presenting raw logs, it groups failures by pattern, turns each failure into a test case scored by a tunable LLM judge, and surfaces recommendations for resolution.
The benchmark numbers are striking: Agent Lens detects critical failures 20 percent more often than the previous baseline and resolves them at one-tenth the cost of using a general-purpose frontier LLM to diagnose traces. For enterprises running agents at scale in healthcare, financial services, and customer operations, that cost differential compounds quickly.
Agent Lens supports deployment as SaaS, in a dedicated cloud, or on-premises for data sovereignty requirements. It includes PII redaction, role-based access so product managers can review conversation-level traces while engineers see raw technical detail, and a direct connection back to Weights and Biases Models so production failures become training signals rather than bug tracker items.
CoreWeave ARIA (Generally Available)
ARIA (AI Research and Iteration Agent) moves from preview to GA with Forge. It is a coding agent that lives inside the development environment and works across the entire improvement loop. Connect it to a GitHub repository and ARIA reads training scripts, agent code, and configurations alongside experiment results. When two branches diverge in behavior, it compares the code rather than just the metrics, drafts the fix on a branch for human review, and launches the next experiment once approved.
For agent builders specifically, ARIA adds tracing and evaluations to an existing repository automatically, and tests prompt and model changes against real production traffic. ARIA is also available in an iOS app, so researchers can monitor runs and interact with their experiments away from a desk.
CoreWeave Sandboxes (Generally Available)
Sandboxes move from preview to GA with Forge. They provide a fresh, isolated CPU or GPU execution environment for every agent tool call, RL training run, or evaluation, removing the need to stand up and tear down infrastructure for each test. Sandboxes are billed per millisecond and run on serverless infrastructure or on dedicated training infrastructure a team already has provisioned.
At Fully Connected, CoreWeave also announced support for NVIDIA Vera CPUs, the first processor architecturally designed for AI agent workloads. A single rack of Vera-equipped nodes supports over 11,000 concurrent sandbox environments, with sandboxes starting more than three times faster than on x86, according to CoreWeave’s own benchmarks.
Weights and Biases Models, Registry, Notebooks, and Post-Training
The W&B Models component handles experiment tracking, evaluation, and versioning, and stays connected to runs as they move from experiment to production. Registry keeps every model checkpoint and agent configuration versioned in open formats. Notebooks brings marimo (an open-source reactive Python notebook project CoreWeave acquired) into the loop alongside experiment data and traces. Serverless SFT and Serverless RL provide post-training services for fine-tuning and reinforcement learning, and Model Distillation is available for compressing larger models into smaller, deployable forms.
CoreWeave Forge Components at a Glance
| Component | Status | What It Does | Key Number |
|---|---|---|---|
| Agent Lens | New (GA) | End-to-end agent tracing and failure analysis | 20% better failure detection, 1/10th resolution cost |
| ARIA | GA (from preview) | Coding agent for experiment analysis and iteration | iOS app, GitHub/HuggingFace/Kaggle integration |
| Sandboxes | GA (from preview) | Isolated environments for tool calls, RL, evals | Per-millisecond billing, Kata VM isolation |
| Weights & Biases Models | Integrated | Experiment tracking, evaluation, versioning | Feeds directly into training loop |
| Serverless SFT | Available | Supervised fine-tuning without cluster management | Open to any framework |
| Serverless RL | Available | Reinforcement learning post-training | RL Rollouts in preview |
| Model Distillation | Available | Compress large models into deployable forms | Connects to Sandboxes and Registry |
| Registry | Integrated | Version control for checkpoints and agent configs | Open formats, any cloud |
| Notebooks | Integrated | marimo-based reactive Python notebooks | Open-source project |
| Dedicated Inference | Available | Custom weights on explicit GPU selection | Full execution transparency |
Pricing: $0 individual, $60/month (teams under 50), custom enterprise.
The Competitive Context
The launch places CoreWeave in direct competition with a fragmented but active set of vendors. Weights and Biases (now integrated), Langfuse, Braintrust, and Honeycomb serve parts of the agent observability market. MLflow and Modal address parts of the deployment loop. What CoreWeave is betting on is that no competitor offers the full stack on compute the teams are already running, with billing in one place and data that never leaves the loop.
That bet is consistent with a pattern visible across the enterprise AI infrastructure market. IBM’s watsonx Orchestrate moved to unify agent governance across AWS Bedrock, Microsoft Foundry, and Google’s platform in October. NVIDIA launched its Open Agent Safety Platform to enforce containment outside the model layer. And OpenAI opened a public beta of its managed Agents API in September.
The common thread: the production-agent infrastructure layer is consolidating from a collection of point tools into integrated platforms. CoreWeave Forge is one of the clearest expressions of that consolidation to date.
Pricing and Availability
CoreWeave Forge is available now. Pricing starts at $0 for individual developers, $60 per month for teams under 50 members, and custom contracts for enterprise. All components are available today except RL Rollouts (in preview), which hot-loads checkpoints directly into a training loop for faster iteration.
MasterClass and Canva are named as early enterprise customers. CoreWeave noted that 4,500 customers, partners, developers, and AI leaders attended Fully Connected 2026 in San Francisco.
What Enterprise AI Teams Should Do Now
For teams already running on CoreWeave infrastructure, the practical implication is immediate: Agent Lens can start tracing production agents today without a separate contract or migration. The connection back to W&B Models and Sandboxes is already there.
For teams evaluating agent infrastructure platforms, Forge raises the bar on what “integrated” means. An observability tool that feeds back into training, a coding agent that reads your codebase and proposes experiments, and sandboxes that start in milliseconds on the same hardware your model runs on are qualitatively different from a tracing tool that lives in a separate tab.
The deeper implication is architectural. Enera’s framework for AI-native enterprise transformation has consistently argued that the organizations building sustainable advantages from AI are the ones treating improvement loops as a core system, not an afterthought. CoreWeave Forge is the most concrete infrastructure expression of that argument available today.
The loop matters as much as the model. The teams that close it fastest will outrun the ones that treat each agent deployment as a static artifact.
Sources: CoreWeave Forge launch announcement (September 30, 2026); CoreWeave Forge blog; CoreWeave ARIA blog; CoreWeave Agent Lens product page; Fully Connected 2026 recap