The Agent Has a Cloud Computer Now

The industry’s preferred AI narrative has matured from ask a question to delegate a small kingdom. OpenAI’s newly announced dots are always-on GPT-6 Astra agents with cloud computers, app connections, feedback loops, and an advertised capacity to work toward goals around the clock. This is the obvious destination for the chatbot: not a better text box, but a junior operator with browser tabs, credentials, and an unnervingly blank calendar.

OpenAI says dots connect to more than 4,000 apps, can be reached through ChatGPT, Slack, or Teams, and may inspect their own cloud desktops while working. Its controls are correspondingly elaborate: proactive research is read-only; sensitive actions get auto-review; users can set rules that allow, require approval for, or block actions; password changes remain human territory. The product announcement also contains the line that ought to be printed above every enterprise procurement desk: “Dots can still make mistakes, so always review consequential work.” Quite. A cloud employee who needs supervision is not a labor arbitrage miracle; it is an operations design problem wearing a demo-video smile.

That tension is the real AI story this week. Capability is becoming ambient before authority becomes legible. The models can increasingly initiate, browse, classify, route, and draft across the organization. The difficult question is no longer whether an agent can click the button. It is who defined the button, what data it saw first, what state it preserved, and which exhausted human is on the hook when it clicks the wrong one.

The Model Context Protocol has completed the standard infrastructure journey: derided by people who had actually tried it, then adopted by those same people once refusing it became more inconvenient than supporting it. Earendil’s Pi now makes MCP core functionality after previously arguing against it. Its stated rationale is less ideological than architectural: the JavaScript-based sandbox needed to make MCP workable also improves the broader product and makes Jev easier to use. “We believe the best way to positively influence something is to embrace it.” Translation: the protocol is now too consequential to heckle from the balcony.

The significant detail is not merely support, but how Pi supports it. It exposes MCP tools to a JavaScript Codemode sandbox that orchestrates tool calls and preserves session state in the transcript rather than the filesystem. It also distinguishes tools available broadly to the model from deferred or Codemode-specific tools. This is a sensible response to the grimy practical reality: dumping every tool definition into context is not an architecture, any more than importing the entire internet is documentation.

Pi’s post also identifies the unresolved problem correctly: composability. Too many servers still return text optimized for token thrift and harness convenience rather than structured data fit for downstream computation. Agents do not become reliable because a vendor has given them 80 tools and a perky system prompt. They become tractable when discovery, schemas, privilege boundaries, state, and failure behavior are designed as first-class concerns. MCP has become the plumbing. Congratulations to everyone involved; the building still needs drains.

There is a quieter shift beneath the agent theatre: mundane decisions are being pulled out of general-purpose prompting and turned into dedicated services. Sebastian Raschka’s assessment of Jev frames the appeal well. The proprietary TypeSafe AI model offers Choice, Noul, and Score APIs for finite decisions and classifications—ticket routing, screening, evaluation, context selection—where one might otherwise spend premium-model tokens asking a frontier system to answer yes or no with theatrical self-confidence.

On Raschka’s reported IMDb test, Jev Choice reached 96.47% accuracy across 25,000 reviews for $0.6492 and roughly 22 minutes. That is interesting, not dispositive. The model is proprietary; its architecture, training process, and data are nonpublic; repeated runs varied; and the author explicitly notes that unknown training overlap with IMDb cannot be ruled out. TypeSafe says “100% of our data is synthetic (but not the type of crap that is just spit out from an LLM obviously)”. An admirably narrow definition of reassurance.

Still, the direction is hard to miss. OpenAI has announced a Decision API in limited preview, described as focusing Luna on user-defined questions with finite answers. Meanwhile, Pi uses Jev in an MCP workflow to classify issue tone. The emerging agent stack is not one majestic model contemplating the universe. It is a cheap decision layer deciding what to retrieve, which tool to load, whether a request is risky, and whether the expensive model need be bothered at all. The most valuable token is increasingly the one your system has the good sense not to generate.

While software firms promise tireless digital colleagues, the physical economy has declined to become a plugin. A Naked Capitalism compilation relays reporting that Oracle issued force-majeure language around the planned Project Jupiter campus in New Mexico, a proposed 2.45-gigawatt site facing setbacks tied to its energy-resource plans. Bloomberg reporting cited in the piece says roughly 20 banks supplied an $18 billion construction loan. The agent may work 24/7; the substation has other commitments.

The article gathers estimates that should make anyone forecasting frictionless AI abundance locate a calculator and an adult beverage. Moody’s, as summarized there, projects 2026 capital spending of $785 billion by six tracked companies, rising to about $1 trillion in 2027. Morgan Stanley is cited as estimating that more than half of GPU servers sold from 2026 through 2028 may lack power connections. These are forecasts, not settled facts, and the cited Brookings author specifically warns that it is premature to call this systemic risk on the scale of previous credit booms. But they describe the constraint with unusual clarity: the bottleneck may be neither model intelligence nor GPU supply, but power, financing structure, permits, and construction schedules—the famously scalable parts of civilization.

This brings the issue full circle. Enterprise AI is becoming a stack of delegated authority: dots acting across applications, MCP connecting the tool surface, classification models policing the decision edges, and data centers underwriting the whole performance with electricity and debt. Each layer is rational in isolation. Together they turn governance from a compliance appendix into the product itself. The autonomous future will not be decided by whichever model writes the most convincing status update. It will be decided by the boring systems that know which tools it can see, which actions it can take, and whether there is actually enough power to run the apology afterward.