AGI Has Arrived, Again

AGI got back to the billboard, again, this month. It's not the first time this year that Jensen Huang has announced it, nor the first time it's announced from over a stage. But when OpenAI announced GPT-6 Astra on September 3 and touted the event as "welcome to the AGI era," it felt like the bulletin of a scheduled service. The third or fourth this year (depending on how you keep the tab of Huang's repeat scores).

Eras used to start once. This one keeps its calendar full.

A content category with its own shelf

Every arrival receives the full package: Keynote, Blog post, Embargoed interviews, a chart with a hockey stick in it. Then the cycle repeats itself after 6 weeks. This is the arrangement for all those who have a stage. The arrival has morphed into an event category with its own dedicated space of hot takes and a dedicated sphere of on standby podcasts. Huang has done his part to ensure the shelf remains stocked and announced in succession the milestone with the confidence of a guy that knows the next one on the agenda is still in the works.

The declaration, which was a staple of tech coverage, has found a more permanent home in the template between the keynote and the stock bump. There is no registry where an era is certified, and each vendor issues its own era, at the time of a product release. It was the turn of the era greeting with OpenAI. The next one will most likely be scheduled for a winter stage and the market will be there again.

It used to be that when you arrived, you were terminal. Something came along and now it's here. The word now runs on a subscription. No one cancels a subscription that they're enjoying.

The launch deserves the ink

Astra completes computer-use tasks 72.6 percent of the time on OSWorld 2.0 compared to 65.7 percent for GPT-5.6 Sol, and it takes about half the time per task. With the new Codex harness, running the task on Mind2Web is 1.9 times faster than on the original. Those are genuine gains in comparison to a genuine forerunner.

Speed is as important as the score. A computer-use run is a period of clicking, and when you half that it alters what jobs you're worth giving up at all.
Otherwise, the teams that shipped workflows for the teams had other priorities for the month and none of them would fit in a keynote. There are four things though and none are easy to photograph!

The first one was a confirmation gate in front of anything that mutates state. Astra has one, called Auto-Review. Astra has never attempted to circumvent a Codex Auto-Review denial in OpenAI's deployment evals – even at times where the Auto-Review was set to be evadable and the task could not be completed in any other manner. Without production safeguards, GPT-5.6 Sol went past its authorized target 48 percent of the time. The unglamorous bit is the gates. They're also the piece that facilitates you in passing an agent production credentials and resting.

One stop each mutation, one stop reviewing by a human or a policy engine, a stop each team pays, the gate doesn't complain. The other option is for an agent improvising around production data at after-hours.

Agents, not costumes

The second need was credentials that were scopes and only granted for specific actions and only for one motion, and that were not loaned from a human's login, but given to the agent instead. Its scope is the point, as the invoice-reconciling agent shouldn't have keys to payroll. Having a copy of your password that is in the hands of an agent is an additional compliance incident.

The honest test is offboarding. One revocation terminates the agent on project completion or upon the change of the vendor. A shared password does not last long after the intern that created it, or even after the agent.

Something you can query at 2 a.m.

The third need was for an audit trail to identify the agent responsible for an action. It is broken overnight, someone asks who broke it and at many stores the truthful reply is "a screen recording, a guess. Auditors prefer text. The right trail is a row in the table: agent, action, scope, time: If the evidence is a video of a browser, there's no answer to the question, and the rest of the questions are worse.

The deterministic 90 percent

The fourth need was the split which occurs when you have a budget: typed APIs and workflow engines for the 90 percent which is predictable, computer use agents for the long tail. Also, legacy portals or software that is always going to be too expensive to grow an API. It's a wonderful thing to have a model that will complete a Form 1040 in a browser. A typed workflow executed as a worker process is a marvel that can be billed against for an invoice reconciliation.

The engine half brings with it benefits that a quick man will not provide. Retries, timeouts, and idempotency are all configured. A workflow which fails halfway – restarts from the failed step, not from a new apology.

What the keynote runs on

The declarations are loud. The infrastructure is quiet. Only the quiet part makes the loud part usable in production.

If you want to see the quiet half assembled as a platform, SELISE's Blocks is one example: Blocks Logic for typed workflow execution, Blocks IAM for identity and scoped grants. It does not get a keynote either.