In Today’s Issue:

🔁 The loop ARIA actually closes, and where its autonomy stops

💸 The guardrails on an agent that can spend your compute budget

🔍 Signal versus noise across thousands of runs

🏗️ Whether the payoff depends on CoreWeave hardware

🔒 How your proprietary experiment data stays yours

🧠 Which models run under the hood (the answer may surprise you)

🔭 The most autonomous ARIA to come, and what stays human

Dear Readers,

Every AI lab runs on a loop that never fully closes. You train a model, then stare at thousands of runs and tens of thousands of metrics, hunting for the one signal that tells you what to try next, and then you start over. Most of that work is manual, and most of it reaches the next experiment too slowly to matter.

On June 29, CoreWeave dropped an agent straight into that loop. ARIA, short for AI Research and Iteration Agent, lives inside Weights & Biases. It reads your experiment data the moment you open a project, forms a hypothesis, launches the next run, and points you toward what to try next. CoreWeave's Chen Goldberg calls it "what the self-improving agent loop looks like in practice" and "a meaningful step on the path to superintelligence."

That is a bold claim for a preview product, so we went to the source. We sent CoreWeave seven pointed questions, and Phil Gurbacki, VP of Product for Weights & Biases, answered every one. His replies are below, in his own words.

All the best,

Kim Isenberg

CoreWeave spent roughly $1.4 billion to acquire Weights & Biases in 2025, and ARIA is the clearest sign yet of what that bet was for. Announced on June 29 and now in public preview, the agent sits inside the same W&B platform that frontier teams already use to track their training runs, and it turns that pile of experiment data into an active research partner. We asked CoreWeave to go past the launch language, and Phil Gurbacki, VP of Product for Weights & Biases, took our seven questions.

Scheduling meetings isn't why you get out of bed. Filing expenses isn't your calling. So hand it all to Catch — the AI admin that takes the tedious work off your plate and gets it done.

Just ask. Catch handles the back-and-forth on scheduling, flights, restaurants, follow-ups, vendors, clients and more — then comes back with it sorted, in moments.

No guesswork. No dropped balls. What's yours stays yours, private by default and reliable by design.

Nobody became a leader to book their own flights. Meet your admin savior and go from swamped to sorted.

Get started free at catchagent.ai — and start spending your time on the things only you can do.

Get Started Now

Q1. Chen Goldberg called ARIA "what the self-improving agent loop looks like in practice" and "a meaningful step on the path to superintelligence." That is a strong claim. Concretely, what is the loop ARIA closes today that a researcher could not already close with an LLM plus the W&B API, and where does the autonomy still stop?

ARIA closes the research iteration cycle by reading experiment results, forming a hypothesis, launching the next run, and recommending where to go from there continuously, with far less manual setup than a researcher would otherwise need. Unlike a general-purpose LLM stitched together with the W&B API, ARIA is purpose-built to understand how experiment data is organized inside W&B, so it knows exactly where to look and what to query rather than scanning indiscriminately across experiment data.

The other key difference is verification: AI research is inherently visual, and ARIA builds the right chart directly inside the platform where the researcher is already working, so findings can be confirmed at a glance rather than pieced together across a chat window and a separate dashboard. The result is a compounding research workflow where each pass turns existing data into the next hypothesis and the next experiment, accelerating model improvement in a way no manual process can match.

Q2. ARIA forms hypotheses, launches experiments, and evaluates results around the clock, and launching real training runs costs real compute. What sits between ARIA and a team's compute budget? Does it need human sign-off before spending, and who is accountable if it burns through a budget chasing a dead-end hypothesis?

ARIA is instructed to ask for permission before incurring any cost. ARIA will be free for a limited time, and once that period ends, researchers will also have a dashboard for full visibility into spending, so there are no surprises.

Q3. When you analyze tens of thousands of metrics across thousands of runs, spurious correlations are almost guaranteed. How does ARIA distinguish a real signal from noise, and what does it give a researcher to verify a recommendation before they act on it?

More data doesn't have to mean more noise. With a larger sample size and more experiments to compare across, scale can actually sharpen the signal rather than bury it, as long as you can analyze it efficiently. And that last part is what ARIA is built for. For example, W&B lets you group runs by parameter value and see each group in aggregate instead of as a clutter of individual runs, and ARIA knows these sorts of tricks for both navigating large amounts of data and visualizing it for verification, so it surfaces patterns that hold up across many runs rather than latching onto a single anomalous result.

Just as important is what the researcher gets to check the recommendation against. Rather than digging through hundreds of thousands of panels by hand, you can ask ARIA to assemble what matters: a section for the most important runs, one for the key panels, and a dedicated section of outliers so nothing slips through. Verification stops being a needle-in-a-haystack search you do manually and becomes something you can see laid out in front of you before you act.

Q4. ARIA expands what CoreWeave describes as a unified agentic AI capability connecting training, inference, and observability. How much of ARIA's value depends on a team running on CoreWeave infrastructure? Does a team that trains elsewhere but logs to W&B get the full agent, or a reduced version?

ARIA works fully for any team logging to W&B, regardless of where they train, so there's no vendor lock-in. That said, ARIA will have access to the full set of infrastructure performance data available through deep integrations with the CoreWeave infrastructure. For example, W&B automatically pulls infrastructure issues from Mission Control, such as GPU errors or thermal violations. The agent itself is the same either way; it just has those infrastructure signals connected to the training picture when they're available, so it can reason across both layers, for instance, tying a sudden slowdown to an infrastructure issue rather than the model.

Q5. ARIA reaches across projects and into teammates' experiments, and frontier labs are protective of exactly this kind of experiment data. What is the data-handling and isolation model, and is anything from a customer's runs used to improve ARIA itself or any underlying model?

Data access in ARIA is strictly scoped to the individual user. Teams can't view each other's experiments unless they already have access, and ARIA operates within a secure sandbox using temporary user credentials. Enterprise and pro customers have additional controls for how their data is used to ensure privacy.

Q6. ARIA is built using W&B Weave as the agent-development platform. What model or models actually power the agent's reasoning, and was that a deliberate choice to stay model-agnostic versus betting on one frontier lab?

ARIA currently runs on OpenAI models, with CoreWeave continuing to test and iterate as the technology evolves. The approach reflects a deliberate focus on choosing the best available model rather than being locked into a single provider.

W&B Weave is used to evaluate ARIA's agent harness, the underlying OpenAI model, and the prompting strategies together as a single agent. This allows us to continually monitor ARIA's performance in production and improve it over time.

Q7. The preview points toward deeper autonomous research capabilities. Looking out a year or two, what does the most autonomous version of ARIA look like, and what is the part of research you think should stay with humans?

Humans stay firmly in the loop, no matter how autonomous ARIA becomes. The agent is built to support pattern recognition, validation, and even hypothesis generation in collaboration with researchers, but humans set the objectives and success criteria ARIA operates within. The vision is a researcher managing ARIA the way they would manage a capable intern, handling the heavy analytical lifting while the researcher steers the direction.

Claude vs Gemini. GPT-7 vs Llama 5. Which AI lab ships AGI first. These are live Kalshi markets with real money on both sides, updated in real time as releases land. The person who follows model cards and tracks evals has a genuine edge here. If that's you, trade it.

Trade the model race

The through-line across Gurbacki's answers is restraint wearing the clothes of ambition. ARIA will read every run and launch the next wave on its own, yet it asks before it spends, scopes each user's data to that user, and treats the researcher as the manager rather than the managed.

The self-improving loop is real, but the human still sets the objective and signs the checks.

ARIA is the first visible payoff from CoreWeave's roughly $1.4 billion bet on Weights & Biases, and it points where MLOps is heading: not dashboards you read, but agents that read them for you and act on what they find. Whether that compounds into genuinely self-improving research or simply faster iteration will come down to how much autonomy teams are willing to hand over. For now ARIA is free in public preview, which is the quickest way to find out for yourself. Our thanks to Phil Gurbacki and the CoreWeave team for the exclusive.

Sources:

🔗 CoreWeave, ARIA launch announcement: https://www.coreweave.com/news/coreweave-aria-launches-as-an-ai-research-and-iteration-agent-with-autonomous-research-and-collaborative-intelligence

🔗 Weights & Biases: https://wandb.ai

🔗 SiliconANGLE, ARIA coverage: https://siliconangle.com/2026/06/29/coreweave-debuts-aria-agent-automate-ai-research-weights-biases/