Claude 4.7 and the Era of the Agentic Office
Anthropic's latest model can run an eight-hour task without supervision. The implications for knowledge work are not subtle.
The most interesting line in Anthropic’s Claude 4.7 release notes was buried in the technical appendix. “Sustained agentic tasks of up to eight hours with periodic checkpointing.” For anyone who has watched the previous generation of agents collapse after forty minutes of autonomous work, that sentence is the entire story.
I have been running Claude 4.7 against a deliberately ambiguous workload for the last two weeks: triaging a backlog of bug reports, drafting follow-ups, and proposing — but not merging — fixes. The previous model could handle the first two reasonably well. This one handles all three, and more importantly, it handles the connective tissue between them. It remembers what it has already done. It does not re-ask questions it has already answered to itself.
The agentic office
There is a category of work that is not creative, not technical, and not strategic, but is nevertheless what most office jobs consist of. Assembling. Reconciling. Following up. Closing loops. This is the work that Claude 4.7 does extremely well now, and it is the work that, for the first time, I can imagine being substantially handed off to a model.
That is not a takedown of office workers. It is a description of what the next two years are going to look like.
The interesting question is not whether models will be able to do agentic work. They can. The interesting question is what humans do with the time we get back, and whether the institutions that employ us let us keep any of it.
What it actually feels like
The thing I did not expect was the personality shift. Claude 4.7 is quieter than 4.6 was. It asks fewer clarifying questions. It is more willing to make a defensible call and proceed, then surface its reasoning at a checkpoint. This is, in my experience, the correct behavior — but it requires you to trust the model in a way that the previous generation did not.
You learn to trust it, or you do not, fairly quickly. After two weeks I am about 80% of the way there. I still review every external action. I am almost never overruling.
Where it still falls down
- Visual reasoning on dense interfaces is still patchy. Anything with a tightly packed dashboard is a coin flip.
- Long-horizon planning under genuine ambiguity — the kind where the goal itself shifts mid-task — remains a place where humans still do better.
- Anything requiring social calibration — knowing when to escalate, when to wait, when to ask a person rather than a system — is the obvious next frontier.
The thing this makes obvious
Tooling around models is now the bottleneck. Claude 4.7 is bottlenecked by its sandbox, by its tools, by the latency of the systems it is asked to operate. The model is not the limiting factor anymore. Make of that what you will.
If you are building anything in this space, the lesson of 4.7 is simple: assume the model will be substantially better in twelve months, and design for the day after that. The companies that win the next cycle are the ones that figure out the interface between agentic models and human institutions. The model itself is, increasingly, a solved problem.
Almost.