Claude 4 arrives, and with it the long-running task
Anthropic released Opus 4 and Sonnet 4 yesterday, emphasising sustained work over long tasks. That capability has a specific consequence for how you supervise it.
Anthropic released Claude Opus 4 and Claude Sonnet 4 on 22 May 2025, with hybrid reasoning and an emphasis on sustained performance across long-running, multi-step work rather than single answers.
The capability being advertised is duration: a model that keeps working coherently on something for a long time. That is genuinely new, and it changes the supervision question in a way worth stating plainly.
Why duration is a different problem
A model that answers in four seconds is easy to supervise. You read the answer. If it is wrong, you notice immediately, before anything has happened.
A model that works for twenty minutes across fifteen steps produces something you did not watch being made. If step four was wrong, steps five through fifteen built on it, and the output looks complete and coherent because it is coherent — it is coherently based on a wrong step four.
This is a genuinely harder review problem, and it is not solved by the output looking good.
What this means for an SME, concretely
Almost no Swiss SME will run twenty-minute autonomous tasks this year. But the capability will appear inside products you buy, described as « it handles the whole process ».
When it does, two questions:
- Can I see the steps? If a long task fails, the log of intermediate steps is the only way to find where. The 11 March post asked this of agent toolkits; it applies doubly here.
- Where does it stop and ask? A fifteen-step process that runs to completion without a checkpoint is fifteen chances to go wrong and one chance to notice. A well-built one pauses at the step that commits something.
The unchanged rule
Longer autonomy does not change the line drawn in December and reinforced on 12 May: preparing is unlimited, completing anything irreversible needs a person.
If anything, duration strengthens the argument. The longer a system works unsupervised, the more important it is that the last step — the one that sends, pays or commits — is still a human clicking something and knowing why.