CoreWeave Launches ARIA, a Coding Agent That Runs the AI Improvement Loop
CoreWeave has made ARIA, the AI Research and Iteration Agent, generally available inside its Forge platform. According to the company, ARIA writes and runs code against a project's experiments, metrics, artifacts and configurations, proposes the next experiment and launches it once the user approves.
- CoreWeave says ARIA, the AI Research and Iteration Agent, is generally available in CoreWeave Forge.
- ARIA analyzes experiments, proposes the next one and launches it after user approval, or runs end to end if configured.
- It works with Weights & Biases Models, Serverless RL, Serverless SFT and Agent Lens.
- ARIA runs in the cloud, keeps project-scoped memory and is also available in an iOS app.
- It connects to GitHub, Hugging Face and Kaggle, and drafts code changes on branches for review.
What ARIA does
CoreWeave frames ARIA around a loop that teams follow after launch: run, observe, curate, improve, evaluate, repeat. The company says that loop does not run itself, because someone has to read results, form a hypothesis, launch an experiment and judge the outcome, work that "has historically required a PhD", in the company's words.
ARIA is meant to work alongside users at each stage. It finds what is driving performance, suggests what to try next and starts the experiment after approval. CoreWeave says this applies whether a team is improving an agent's prompts against production traffic or training a model down to the weights.
The company describes a typical day. An experiment finishes, and an Automation set up in plain language fires. ARIA reviews the configuration, metrics and training behavior, then writes a Report on what changed and why. Users can open the chat from a phone, ask what ARIA would try next, approve a drafted change on a branch in a connected repository, and see the next experiment launch on their own infrastructure.
According to CoreWeave, this works with training models using Weights & Biases Models, post-training with Serverless RL or Serverless SFT, and building agents with Agent Lens. For model trainers it covers training jobs from pre-training to fine-tuning and RL. Users choose how much control to hand over: ARIA can pause for approval before each launch, or run end to end with results checked at the end.
Visual analysis, cloud execution and shared memory
ARIA lives in the Forge dashboard and answers with live visualizations rendered in the chat. Standard charts handle common questions. For other needs, such as distributions, heatmaps and scaling laws, CoreWeave says ARIA builds custom interactive plots on the fly. When a finding is worth keeping, it can create a persistent Report or dashboard view that updates as new experiments arrive and can be reviewed by the whole team.
The agent runs in the cloud rather than on a user's machine, so a deep analysis across hundreds of experiments continues after the user steps away. Automations can fire on events such as an experiment completing or a metric crossing a threshold. If configured, ARIA can draft the Report or launch the follow-up experiment on its own. An iOS app lets users check on training, read findings and ask for the next step remotely.
Memory is scoped to the project, not to one person's session. CoreWeave says ARIA remembers what experiments showed and what the team decided, so every conversation starts from shared learnings. The same memory is meant to keep managers and stakeholders informed, with a question about the team's week answered alongside a concise visual summary of the experiment data.
Integrations and availability
With GitHub connected, ARIA reads training scripts, agent code and configurations alongside results. CoreWeave says that when two branches behave differently, it compares the code and not only the metrics, and drafts each change on a branch for review. For agent builders, pointing ARIA at a repository lets it add tracing and evaluations. Connecting Hugging Face or Kaggle lets it pull the models and datasets that experiments depend on.
CoreWeave ARIA is generally available today as part of CoreWeave Forge. Users can open it from the ARIA home page, look for Ask ARIA inside Forge, or start a conversation in the iOS app. The source does not give pricing details.
For teams already working in Forge, the release puts experiment analysis, follow-up planning and launches in one assistant, with human approval as an optional gate. How much of the loop teams choose to automate is left to them, according to CoreWeave.
