GPT-6 Astra and Claude Fable 5.1 arrived this week. To me, they point to a change in what businesses can ask AI to take on. More demanding office tasks can now be given to an agent that carries out the work. That puts a greater responsibility on us to define what it is allowed to do.
A lot has changed since May
In May, we wrote about using AI in our development work. Much of that story was about help with writing code, preparing drafts and running tests. It now describes a smaller share of the work we can ask an agent to handle.
OpenAI released GPT-6 Astra on September 3, combining reasoning, coding, computer use and document work. It is available in products including Codex and ChatGPT Work. Anthropic released Claude Fable 5.1 on September 1, focused on demanding coding, knowledge work and extended tasks.
Computer use was already available in May. What interests me is how much of a connected task we can now try handing over: investigating the problem, making a plan, building something, testing it and using the results to correct the work.
There are concrete examples. OpenAI's account of building Void Explorer describes using Codex for code, models and testing, with people judging the game and directing its development. In Anthropic's launch announcement, Ramp describes an unattended Fable 5.1 run lasting 38 hours. That is a customer account published by the vendor, not a promise that every task will work the same way.
These examples show agents handling long tasks with many steps. How long they can work without help still depends on the task, the tools and the access they are given.
Ask for the work you need
Need a clearer view of your business in Power BI? Ask for a report showing where margins are falling, which customers are growing and what deserves closer investigation. Microsoft's Copilot can create and edit report pages from a description. That does not give Astra automatic access to Power BI. The tool needs the right data, and someone needs to check that the figures support the report's conclusions.
Want to harden your company's systems? Ask for a review of code, permissions and weaknesses, with changes that can be tested. Codex Security includes a workflow for proposing security hardening. I would start with read access and a test environment, and require approval before changing production.
Need a business app? Describe its users, the job they need to do and the information they need. Have AI help build a working prototype. Before putting it into use, review authentication, storage and the awkward cases that a quick demonstration tends to miss.
Does your working day involve five systems and a spreadsheet? Ask how they could be connected through one interface designed around your workflow. You may be able to keep the existing systems and make them easier to use. Accounting requirements, history and access rights still matter when the new interface looks simpler.
I expect some office jobs to disappear
I think many office jobs will change quickly, and some will disappear. Tasks built around gathering information, comparing it, moving it between systems and writing a report are obvious candidates.
That is my assessment, not a documented timetable for the labour market. A job also involves responsibility, priorities, relationships and knowledge that is rarely collected in one document. A company's ability to change depends on more than the model.
Even so, I think businesses should start investigating this now. Choose one real task. Give the agent a bounded environment. Compare its results with the current process, including mistakes and the time people spend checking its work.
Developers need to understand the whole system
It is tempting to conclude that better AI reduces the need for technical knowledge. My conclusion is that the developer's responsibility gets broader.
When an agent can produce a lot of code, understanding whether the solution is well designed becomes more valuable. Where should data live? Who can read and change it? What happens when an integration fails halfway through? Can we trace a mistake and undo the change?
We still need to read code. We also need to assess architecture, security, operations and the connections between systems. We need to understand the limits of our AI tools and design tests that can expose a misunderstanding. A passing test tells us little if it simply repeats the mistake in the implementation.
Set boundaries before assigning the task
To me, the next stage is increasingly about clear goals and limits. An agent should know which systems it can use, what it may change and when to stop and ask. OpenAI's computer use guidance includes isolated environments, limited access and human review for actions with significant consequences.
“Make the business more efficient” is a poor assignment. “Draft a weekly report using these three data sources, without changing them or sending anything to anyone” gives us something we can check.
These tools give us room to be more ambitious. I want to use that capability to build better software, and spend more of the developer's time on the decisions that make it safe and useful.
Have a task you would like to explore? Tell us what it involves. That is a better starting point than choosing a model.