CoreWeave Forge links the full AI improvement loop in one environment

A connected loop instead of a patchwork of tools

According to CoreWeave, Forge covers serving models in production, observing in-production conversations, curating data, building the next version and evaluating the final product before it ships. The company says many teams already run this cycle across disconnected systems and manual workflows, and that they often stall in the handoffs between tools.

CoreWeave describes the problem in concrete terms. Once a model is in production, AI engineers who know what good output looks like hand it to an application or SRE team. Application traces land in an observability dashboard built for uptime, which shows that a service is healthy but says nothing about answer quality or user sentiment. Evaluation suites drift because nobody on the production side owns updating them, and researchers do not see the traffic.

Forge is meant to reduce those manual handoffs. CoreWeave says it stays open to whichever model, framework or cloud a team already runs.

Five stages, each with its own tools

The company frames the cycle as run, observe, curate, improve and evaluate. For each stage it names a specific challenge, from capturing the right signals to defining improvements against clear standards rather than subjective judgment.

Several components are new or newly available:

• CoreWeave Agent Lens (new service) traces every step, decision and tool call end to end. Monitors score live traffic with human oversight so prioritized failures surface early, with recommendations. CoreWeave claims it improves failure detection by 20% and fixes issues at half of the cost.
• CoreWeave ARIA (now generally available) helps users analyze runs, propose experiments and recommend code changes, which it stores in GitHub.
• CoreWeave Sandboxes (now generally available) runs agents, tool calls, RL and evals in isolated CPU or GPU environments.
• Model Distillation (new service) trains a smaller open-weights model on a larger model's outputs for a task already proven in production, then scores it head-to-head against the current model.
• CoreWeave Notebooks (new service) offers managed Python notebooks for sharing evals and custom analytics per model and per agent.

Inference, training and a system of record

Forge also includes CoreWeave Inference, which offers leading open-weights models through Serverless Inference or Dedicated Inference on isolated infrastructure for larger workloads. Weights & Biases Models covers experiment tracking, hyperparameter sweeps and automated workflows, and is integrated with ARIA to support autoresearch, described as autonomously improving a model based on detected signals.

Post-Training uses a team's own production signal to improve model quality and cut latency and cost, with no training cluster required. Serverless SFT and Serverless RL let users experiment with their own training recipes.

CoreWeave Registry acts as a single source of truth for models, agents and datasets. According to CoreWeave, every dataset, evaluation, model version, trace and deployment stays connected, so teams can see what changed, why, and whether it improved results.

What it means for AI teams

CoreWeave's pitch is that a tighter loop shortens the time from a production issue to a validated fix with every cycle, or in its words, "the more you ship, the faster you improve." For teams that spend more time maintaining tools and coordinating handoffs than improving their applications, Forge is positioned as a way to shift that balance. The performance figures for Agent Lens are the company's own claims, and the source does not detail how they were measured.