CoreWeave Launches Forge to Connect AI Deployment, Monitoring and Improvement

CoreWeave Releases ARIA to Analyze AI Experiments and Run the Next Round

CoreWeave announced general availability of ARIA on October 1, 2026, putting an AI research assistant inside its Forge platform. The company says the agent can analyze experiment results, propose changes and launch another run. Users can require approval at each launch or configure it to carry the experiment cycle through without stopping.

ARIA stands for AI Research and Iteration Agent. According to CoreWeave, it writes and runs code against a project’s experiments, metrics, saved outputs and configurations. Its job extends beyond summarizing a result: it forms a hypothesis about what to change, proposes the next experiment and helps execute it. For agent builders, that can mean testing prompt or model changes against real traffic.

For model trainers, CoreWeave describes jobs running on their infrastructure, including pre-training, fine-tuning and reinforcement learning—training guided by feedback. Users can set up an Automation in plain language to respond when an experiment finishes or a metric crosses a threshold. ARIA then reviews the configuration, measurements and training behavior, and can write a report explaining what changed.

The execution controls deserve a distinction. CoreWeave’s product page emphasizes approval for each launch and human review before code changes ship. Its launch announcement describes a broader option: users can let ARIA run the experiment cycle end to end and inspect the results afterward.

GitHub: ARIA reads connected code and configurations alongside results, compares branches and drafts changes on a branch for review.

Hugging Face and Kaggle: ARIA can pull the models and datasets used by experiments through connected accounts.

Memory is shared at the project level, rather than confined to one person’s conversation. CoreWeave says ARIA retains experiment findings and team decisions for later sessions. It runs in the cloud, so long analyses can continue after a user closes their laptop; conversations are accessible from Forge or the iOS app.

General availability does not cover every connection CoreWeave describes. The product page marks ARIA’s planned link to Mission Control through MCP, a tool-connection protocol, as coming soon. Its ability to find patterns in team-flagged traces and propose fixes in Forge’s curation stage is also forthcoming, distinct from the experiment analysis available now.

CoreWeave says AI search training raised accuracy while cutting tool calls by 30.24%

CoreWeave says a joint AI search-training test with NVIDIA and you.com raised BrowseComp accuracy from 36.97% to 45.45%, while reducing average tool calls by 30.24%. The October 5, 2026 disclosure details results from capabilities introduced at Fully Connected 2026, including infrastructure that loads updated models into live deployments during training.

The workflows sit within CoreWeave Forge, the development environment announced September 30. Model Distillation, RL Rollouts and a programmable training API connect production feedback to training and back to serving users.

In reinforcement learning, the model generates attempts—called rollouts—that feed the next training step. Updated weights then return to serving. RL Rollouts lets teams retain their trainer, reward logic and environment while CoreWeave handles inference, using NVIDIA Dynamo with vLLM support.

CoreWeave says new checkpoints can load without interrupting requests already underway. After the first upload, teams can send only changed weights rather than full snapshots. The service also returns training metadata, including token IDs and the probabilities assigned to generated tokens.

The joint validation post-trained NVIDIA Nemotron 3.5 Lightning using supervised and reinforcement learning with you.com’s search tools. Checkpoint loading was approximately 15 times faster than traditional redeployment, CoreWeave says. But it credits the workload gains to better search tooling, model refinement and surrounding software changes—not rollout infrastructure alone.

Model Distillation is the managed route: it trains a smaller model to reproduce a stronger model’s behavior. Agent Lens captures production traffic for training. Teams can relabel weak answers, test models and training settings, then compare candidates against the model already in production.

CoreWeave points to Method’s model for navigating automated bank phone systems. In August 2026, Method refreshed production data and relabeled examples with GPT-5.6 Sol. The resulting model won 57.5% of head-to-head evaluations against its production predecessor, CoreWeave says. The base model stayed the same: Llama 3.1 8B.

The programmable API offers more control: teams write Python training loops while CoreWeave manages the infrastructure. They can mix supervised learning from example answers with reinforcement learning that rewards successful outcomes. The API remains in limited preview, with an official launch planned for later in 2026.

CoreWeave Opens Agent Lens Preview to Find AI-Agent Failures and Check Fixes

Teams running AI agents can now try a tool designed to turn live failures into tests for the next version, rather than investigate each conversation separately. CoreWeave opened Agent Lens in public preview on September 30, 2026. Its launch announcement describes a connected workflow for finding recurring problems, diagnosing them and checking proposed fixes before release.

Agent Lens works from traces: records of an agent’s inputs, outputs, model responses and tool calls. Its conversation view presents those records as readable dialogue. Product managers and subject-matter experts can review what happened, while engineers can expand individual turns to inspect the execution details behind them.

CoreWeave says the system analyzes the full production collection rather than a small, manually selected sample. It groups conversations with the same underlying failure, even when their wording, tool sequence or visible error differs. Each group links back to source traces and representative examples, giving teams evidence to inspect rather than just a changed dashboard metric.

Those groups update as new traces arrive. Teams can rank problems by frequency, changes over time and affected workflows. Agent Lens also groups requests by user intent, showing which tasks customers want and where the agent falls short. CoreWeave presents the two views as complementary: a popular task may deserve investment, while a rarer failure may demand attention because of safety or compliance risk.

Finding a suspicious pattern is only the start. Agent Lens lets teams create custom LLM judges—language models used to score agent behavior—based on failures found in production. Domain experts define the expected response and score representative examples, establishing the standards the automated judge should apply.

The tool compares human labels with judge scores and surfaces disagreements. Teams can examine cases the judge wrongly flags or misses, clarify scoring rules and refine its criteria before applying it broadly. Once aligned with those standards, the judge can score new production traces automatically. The workflow therefore includes checking the evaluator, not simply accepting its first verdict.

A candidate fix can change a prompt, model, tool configuration or code. Teams first compare its response with the original production case, then run it against the full evaluation set. That second step is intended to catch regressions elsewhere, rather than treating success on one replay as enough to ship.