The news
CoreWeave (CRWV) on September 30, 2026 launched CoreWeave Forge, a development layer meant to help teams take what their AI models and agents do in production and use it to build better versions. The move extends the company beyond its core business of renting GPU capacity.
According to CoreWeave's announcement, Forge is built around a repeating loop: run a model or agent, observe how it behaves, curate the failures into datasets, improve the system and evaluate the next version before it ships. CoreWeave describes the workflow as five stages: run, observe, curate, improve and evaluate.
The platform pulls together tools CoreWeave already offers, including Weights & Biases Models, CoreWeave Registry, Notebooks, Sandboxes, Post-Training and Inference, according to the company's announcement.
One new piece is Agent Lens, an observability system that records traces of what AI agents actually do. It is in public preview on CoreWeave's hosted cloud, with dedicated and on-premises versions planned after the preview, according to AICybr. The same report says Agent Lens insights and online LLM judges, automated graders that score model output, are free through the end of 2026, and a 30-day trial of Forge Pro is available.
CoreWeave said Forge comes in Free, Pro and Enterprise editions and is open across the models, frameworks and other clouds teams use, rather than tied only to CoreWeave infrastructure. The company named MasterClass and Canva as early users building on the platform.
Chen Goldberg, CoreWeave's executive vice president of product and engineering, said in the announcement that engineering teams need to understand the gaps in their systems and improve the next version.
The numbers
- Launch date
- September 30, 2026
- Stages in the Forge loop
- 5 (run, observe, curate, improve, evaluate)
- Forge Pro trial
- 30 days
- Free Agent Lens insights and online judges
- Through end of 2026
Why CEOs should care
For AI buyers, Forge is another option in a crowded field of tools that monitor and tune models, a market where Amazon, Microsoft and Google bundle similar services into their clouds. The open, multi-cloud pitch is the key claim to test. Ask CoreWeave which parts of Forge work fully when your inference runs elsewhere, and what data leaves your environment when traces are collected.
CFOs should look at the pricing path, not just the free period. Agent Lens insights and online judges are free through the end of 2026, which means costs could change in 2027 after teams have built workflows around them. Get the post-promotion pricing in writing before committing production traffic.
CISOs and risk teams should treat agent traces as sensitive data. Production traces can hold customer prompts, internal documents and tool outputs. Before turning on tracing, confirm retention periods, access controls and whether dedicated or on-premises Agent Lens will arrive on a timeline that fits your compliance needs.
The bigger picture
CoreWeave built its business renting Nvidia GPUs to AI labs and enterprises. Forge signals a push into higher-level software, the layer where customers tend to stay longer. Constellation Research analyst Holger Mueller said the AI cloud market is entering a software platformization phase. If that holds, specialist GPU clouds will increasingly compete on tooling, not only on price and chip supply.
What’s next
Watch for general availability of Agent Lens, the arrival of its dedicated and on-premises versions, and the pricing CoreWeave sets once the free period ends after 2026. Customer adoption beyond the named early users will show whether enterprises want this layer from their GPU provider.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
