GPT-6 Astra Ships a Document-Writing Computer-Use Agent Right as China Codifies Who Gets to Approve One

Three stories landed within a few weeks of each other, and none of them is really about a model beating a benchmark. OpenAI shipped a chatbot that can fill out your spreadsheets and click through your browser. China wrote a law that sorts every agent decision into who is allowed to make it. And Gartner's latest numbers show the jump from "agents as a pilot" to "agents as a checkbox in the RFP" is happening faster than most roadmaps assumed. Put together, they're the same question asked three different ways: now that a chatbot can act instead of just answer, who has to say yes first?

1. GPT-6 Astra Turns ChatGPT Into a Computer-Use Agent

OpenAI's GPT-6 Astra began rolling out this week, first to enterprise customers through OpenAI's gated Daybreak access program, with ChatGPT Plus, Pro, Business, and Enterprise users — plus the API and cloud platforms including AWS Bedrock and Azure — following in the coming days. The headline capability isn't a smarter chat reply; it's Astra completing multi-step computer tasks end to end: producing documents, spreadsheets, and presentations from a prompt, filling out web forms, updating CRM records, building and testing a website, or researching a topic and drafting the results straight into a document or email. On OSWorld 2.0's offline benchmark, Astra matched or beat the prior model's computer-use accuracy in roughly 47% less time per task, and it undercut GPT-5.6 Sol's per-task API cost on coding benchmarks by more than half at comparable quality. The model also ships with a 1.05-million-token context window — and, notably, OpenAI is gating its sharpest capabilities behind the same "critical" cyber-risk vetting tier it introduced for GPT-5.6, rather than handing computer control to every account on day one.

2. China Makes Agent Authority a Legal Category, Not a Product Choice

While Astra was rolling out under an internal review gate, China's own answer to the same problem has already been law for nearly two months. The Cyberspace Administration, the National Development and Reform Commission, and the Ministry of Industry and Information Technology jointly issued Implementation Opinions on intelligent agents that became enforceable on July 15. The rule requires every agent's possible decisions to be sorted, before deployment, into one of three authority tiers: human-only, user-approval-required, or agent-autonomous, with the tier scaled to the consequence of the action. Agents operating in healthcare, transportation, media, and public safety face mandatory filing on top of that classification. It's the first time a national regulator has treated "agentic AI" as its own governed category rather than an application layer sitting on top of a chatbot — and it formalizes, as statute, almost exactly the kind of internal gating OpenAI just built into Astra voluntarily.

3. The Adoption Curve Just Went Vertical

The regulatory and product stories are catching up to where deployment already is. Gartner now projects that 40% of enterprise applications will ship with a task-specific AI agent embedded by the end of 2026 — up from under 5% at the start of 2025 — and separate surveys put overall organizational AI adoption at 88%, up from 78% a year earlier. That's no longer an early-adopter curve; it's close to a default expectation for new enterprise software. It also explains why the permission question in stories one and two matters commercially, not just academically: an agent wired into CRM records, forms, and calendars at 40% of enterprise apps is an agent that needs a defensible answer to "who approved this write" long before an auditor or a regulator asks.

Building the Permission Layer Now, Not After

If a chatbot in your product can already draft an email, the temptation is to let it also send one. Astra's cyber-tier gating and China's three-tier statute both land on the same practical shape: classify the action before you let the model reach it, not after something goes wrong.

type AuthorityTier = 'human-only' | 'approval-required' | 'agent-autonomous';

interface AgentAction {
  name: string;
  tier: AuthorityTier;
  execute: (params: Record<string, unknown>) => Promise<unknown>;
}

async function dispatch(action: AgentAction, params: Record<string, unknown>) {
  if (action.tier === 'human-only') {
    return { status: 'blocked', reason: `${action.name} cannot be performed by an agent` };
  }
  if (action.tier === 'approval-required') {
    return { status: 'pending', action: action.name, params }; // surface to a human before execute() runs
  }
  return { status: 'executed', result: await action.execute(params) };
}

The tier lives with the action definition, not the model call, which is what lets you tighten or loosen it as a policy change rather than a prompt-engineering exercise — and what lets you show a regulator, or a customer, exactly which actions your chatbot was allowed to take on its own.

Suggested visuals: a side-by-side timeline of Astra's Daybreak rollout against China's July 15 enforcement date; a three-lane diagram of the human-only / approval-required / agent-autonomous tiers with example actions in each lane; and a bar chart showing enterprise agent-embedded-app share moving from under 5% (2025) to a projected 40% (end of 2026).

— Maya