The most important word in today’s AI brief is not “model.” It is “harness.” OpenAI has put the machinery behind its long-running Codex agents into a managed public API, which means developers can now rent orchestration, context management and execution—not just intelligence by the token. That is the headline in AI news today, and it shifts the competitive question from “Which model scores highest?” to “Which system can finish the work reliably?” The rest of the day tells the same story from different angles. Oracle’s cloud numbers show how much capital is chasing inference demand. Nvidia is opening more of its rack architecture to specialist chips, while Skild says a robot can learn an unfamiliar job from one video. NASA and IBM are turning decades of lunar observations into an open model. And Miro’s sale shows that yesterday’s software valuations still have gravity. Here is what changed on September 10—and what deserves your attention next.

AI news today: OpenAI turns the Codex harness into an API

OpenAI launched the Agents API in public beta for all developers on September 10. It packages the open-source Codex harness with managed infrastructure for jobs that can run for days, work with files, execute code and save intermediate results. Automatic context compaction, tool search, programmatic tool calling and multi-agent coordination are built in. Developers can use an OpenAI-hosted sandbox, their own infrastructure or partner environments; there is no separate API fee beyond model tokens and tools.

Why it matters: the difficult part of production agents is increasingly everything around the model—state, tools, retries, observability and safe execution. OpenAI is turning that layer into a product. The customer performance figures in its announcement are testimonials, not independent benchmarks, so builders should test reliability, cost and failure recovery on their own workloads. Watch the beta’s path to general availability, especially its controls for long-lived tasks and third-party sandboxes.

Oracle’s AI backlog meets a 121% infrastructure surge

Oracle reported record fiscal first-quarter revenue of $19.3 billion, up 30% year over year. Cloud revenue rose 62% to $11.6 billion, while cloud infrastructure revenue climbed 121% to $7.4 billion. The company delivered another 850 megawatts of data-center capacity and ended the quarter with $664 billion in remaining performance obligations.

That $664 billion is Oracle’s total contracted backlog, not a pure AI number—a distinction worth keeping when the zeros get theatrical. Even so, management said it booked more than $30 billion of additional AI cloud contracts during the quarter and that demand still exceeds supply. For buyers, the next question is whether Oracle can convert contracted work into live capacity without letting power, construction or chips become the rate limiter. This continues the infrastructure race we tracked in our September 3 cloud brief.

NASA and IBM open a foundation model for the Moon

NASA and IBM released an open-source Lunar Foundation Model trained on a new machine-learning-ready dataset with more than 30 spatially aligned layers from nine instruments across four missions. Researchers can adapt it to look for potential ice deposits, map craters and study volcanic features without building a separate model from scratch for every question.

An IBM- and NASA-authored technical paper reports up to a 22% reduction in error when identifying high-potential lunar-ice areas versus SwinV2-B, plus nearly 19% better context-scale crater mapping with half the training data. Those are promising research-team benchmarks, not an independent replication. The larger consequence is more durable: scientific agencies are beginning to publish reusable models and curated data together, giving outside researchers something concrete to inspect, fine-tune and challenge.

Nvidia gives d-Matrix a faster route into AI racks

Inference-chip startup d-Matrix said its next-generation Raptor XPUs will connect to Nvidia’s platform through NVLink Fusion, MGX rack designs and Spectrum-X networking. In plain terms, d-Matrix can focus on its accelerator while borrowing a mature ecosystem for networking, power, cooling and deployment. Financial terms were not disclosed.

This is both openness and leverage. Nvidia makes its rack architecture more attractive by welcoming non-Nvidia processors, while specialist chipmakers avoid rebuilding an entire data-center stack. Customers get more choice inside a common architecture—but Nvidia remains the connective tissue. The announcement is a planned integration, not a shipping product; the practical test is whether Raptor reaches production systems on schedule and wins workloads on cost per token, not presentation-slide compatibility.

Skild says one video can teach a robot a new task

Skild AI detailed its S1 robotic foundation model, designed to infer a previously unseen, multi-step task from a single video demonstration without task-specific retraining. Nvidia says the model uses its infrastructure, Isaac simulation tools and Cosmos data tooling. Demonstrated tasks include plant potting, coffee brewing and assembly sequences lasting as long as ten minutes.

Skild reports roughly 66% success per step on new multi-step tasks, versus 9% for a comparable system, and says one video can provide value similar to about 380 physical demonstrations. Treat those numbers as company-reported until methods and results are independently reproduced. Still, the direction matters for factories: programming by demonstration could make robots useful in environments where jobs and layouts change too quickly for bespoke automation.

Miro’s $1.355 billion sale resets the collaboration-software clock

Bending Spoons signed a definitive agreement to acquire Miro at a $1.355 billion enterprise value, implying about $1.79 billion in equity value after net cash. The buyer says Miro has around $600 million in annual recurring revenue, nearly 90% from business and enterprise customers. The all-cash deal is expected to close in the fourth quarter, subject to approvals; the companies remain independent until then.

The price is the sting. TechCrunch notes Miro was valued at $17.5 billion in 2021. That comparison crosses different market conditions and deal structures, but it captures the reset facing once-celebrated SaaS companies. Customers should watch product investment, pricing and staffing after closing; boards should notice that recurring revenue alone no longer guarantees a premium multiple when AI is rewriting the collaboration layer.

Watch & Learn

Editor’s note: In 13 minutes, practitioner James Briggs shows how the OpenAI Agents SDK connects to the platform’s Traces dashboard for out-of-the-box observability. It is a practical primer for developers who can already run a basic agent and now need to inspect tool calls, handoffs and failures—the unglamorous work that decides whether an agent belongs in production.

AI, Translated: agent harness

An agent harness is the software layer that keeps an AI agent working through a real task. It manages context, gives the model tools, records state, handles retries and decides when to delegate or stop. Picture a research agent that searches the web, saves notes, runs a calculation and asks another agent to check the answer: the harness coordinates that sequence. You should care because a strong model inside a weak harness can still lose its place, repeat work or fail silently.

Try This Today: build an agent acceptance-test project in Claude

Goal: turn one planned AI workflow into a reusable test before anyone automates it.

1. Create a Claude Project called “Agent Acceptance Tests.” Anthropic says Projects are available to all users, with free accounts limited to five.
2. Add the workflow description, allowed tools and three examples of a correct result to project knowledge.
3. Start a new project chat and paste the prompt below. Review the tests before building the agent.

Act as a skeptical QA lead. Turn this workflow into 10 acceptance tests covering success, missing inputs, wrong tool choice, stale sources, partial failure, unsafe action, cost overrun and recovery. For each test, give the setup, expected behavior, pass condition and evidence to log. Flag any requirement that cannot be tested objectively.

You will finish with a compact evaluation set in about ten minutes—and a useful reminder that “it worked in the demo” is not a test strategy.

One thing to remember

The next phase of AI competition is moving outward from the model: into harnesses, infrastructure, data and deployment. The winners will not merely generate the best answer; they will complete the job and leave evidence that they did it properly.