Preface Executive Briefing

Week of 5 September 2026

GPT-6 Astra: When AGI Starts Doing the Work

AI is moving from producing answers to completing work inside the software businesses already use.

Computer use opens a third route into enterprise systems. As agents become more capable, authority, verification and recovery become the advantage.

Astra can now work inside the software people already use

In OpenAI's launch film, Astra works through forms, image tools, documents and slides. The interface is no longer where automation stops.

OpenAI's launch film shows Astra navigating visible software across several multi-step tasks.

Watch the original OpenAI film
Official OpenAI demonstration
72.6%
OSWorld 2.0
Multi-step computer workflows, compared with 65.7% for GPT-5.6 Sol.
47% less time
In OpenAI's latency simulation
About 40 minutes per task, compared with roughly 75 minutes.
92.7%
ScreenSpot-Pro
Locating interface elements across professional software and operating systems.

The strategic signal is the lower cost of testing work for delegation, before every application exposes a perfect connector.

Computer use creates a third route into enterprise systems

A business can now reach the same completed task through a person, a structured software connection or an agent operating the visible application.

Manual execution

A person operates the application

Useful for exceptions, judgement and work that is still changing.

API automation

Software calls a structured interface

Best for predictable, high-volume and high-consequence actions.

Interface execution

An agent operates the visible software

Reaches legacy systems, portals and long-tail workflows sooner.

Same destination

Completed work

The route changes. The business outcome and the need to verify it do not.

Computer use lowers the cost of testing whether work can be delegated. APIs remain the stronger route where reliability, scale or consequence demands them.

Agents are learning to improve the way they work

Intelligence is expanding through larger models, more reasoning time and systems that reuse what they learn across repeated work.

Context

Retain the brief

Keep the goal, constraints and work already completed in view.

Memory

Carry lessons forward

Recover useful decisions and results from earlier attempts.

Skills

Reuse proven methods

Apply instructions and patterns that worked on similar jobs.

Tools

Reach the work

Use software, files and services instead of stopping at an answer.

Feedback

Check and adapt

Inspect the result, repair failures and improve the next attempt.

This is recursion in practical terms: the system keeps useful context, applies proven methods, checks the result and begins the next attempt with more than it had before.

Astra brings the AGI question closer to real work

Astra is not AGI today. It closes part of the distance because intelligence can now reach into the software where work happens. Its ARC-AGI-3 results also show that capability lives in the model and the execution system around it.

Delegated work today

Astra can reason and act through software across bounded professional tasks with a clear goal and operating environment.

The broader AGI threshold

High autonomy and performance across most economically valuable work, extending far beyond selected tasks, environments or benchmarks.

More capable agents need clearer authority boundaries

Observing, preparing, acting and committing carry different consequences. The approval rule should strengthen as an action becomes harder to reverse.

Read the state

Inspect a record, page or document without changing it.

Draft the change

Fill the fields or prepare the transaction for review.

Make a reversible change

Update a record where the previous state can be restored.

Affect another party

Submit, publish or communicate outside the organisation.

Make it irreversible

Release money, merge to production or accept a binding term.

Human confirmation begins here

Verification should inspect system state, while recovery preserves partial work and returns control safely when the environment changes.

The first delegated jobs should be bounded and observable

Four questions reveal whether a process is ready for an agent or still needs to be redesigned.

A

Bounded

Does the job have a clear start, finish and definition of done?

B

Checkable

Can another rule, system or person verify the completed result?

C

Reversible

Can a mistaken action be stopped, undone or recovered safely?

D

Worth the overhead

Does the saved work justify the review, monitoring and recovery cost?

The decision

Delegate with controls

Start with a narrow work package, limited authority, an independent check and a clear handback point.

Good starts: recurring reports, record updates and document-to-form transfer

Redesign the process first

Clarify the finish, expose the state or add a recovery route before giving the agent permission to act.

Hold: ambiguous ownership, hidden state or irreversible execution

The first wave should prove reliable completed work within narrow authority boundaries.

The operating model in four moves

Judge the system by cost per correct, authorised, verified and recoverable outcome.

01

Choose the work

Start with a completed outcome. Prefer work that is bounded, observable and repeated often enough to matter.

02

Set the authority

Separate observing, preparing, acting and committing. Give the agent only the permission required for the next safe step.

03

Verify the state

Check what changed inside the system. The agent's final message is insufficient evidence. Preserve the request, actions and before-and-after state.

04

Design the recovery

Define when the agent stops, what partial work it keeps and how control returns to a person without losing the thread.

FAQ

No. Astra is a major step towards systems that can reason and act across economically useful work, but AGI is the broader threshold of high autonomy and performance across most economically valuable work. Astra brings that question closer to operations without settling it.