When Anthropic released a few new agent skills aimed at doing legal work, a whopping $285B was wiped off the market, with some legal software companies losing as much as 18% of their value.
As simple as it is genius
So what are agent skills exactly? They're Markdown documents. Plain text files with instructions that tell an LLM how to perform a specific job. No new model, no fine-tuning. Just a well-structured document that describes the domain and shows what good output looks like.
If you've ever written a detailed onboarding guide for a new colleague, you've basically written an agent skill. The difference is that the LLM will follow these instructions precisely, every single time, and it can do it at a scale and speed that no human can match.
What Anthropic released were skills for legal work: contract review, compliance checking, document analysis. The thing that should make you pay attention is that anyone with domain expertise can write these. You don't need to be an ML engineer or a programmer. You just need to understand the work well enough to write it down clearly. And that's what really freaked out the markets: the barrier to automating specialized knowledge work just dropped to basically zero.
A huge number of jobs are not that complex
Here's the thing that I think a lot of people in white-collar jobs don't want to hear: much of what makes their work valuable isn't the difficulty. It's the knowledge barrier. Lawyers spend years learning rules and procedures that are, when you really look at them, just information. The work is expensive because few people have access to that information, not because it requires some rare talent to apply it.
All it takes is one person who knows the trade to write it into a skill document, and the gatekeeping is over. That internal knowledge that used to justify high salaries is now something anyone can hand to an LLM. The small empires built on "I know this and you don't" turn out to be pretty fragile once you realize the only thing protecting them was the fact that nobody had written it down yet.
I started my career as a software developer at Switzerland's biggest supermarket chain. As a part of preparation for emergency situations, such as a big earthquake, the company had to take measurements to guarantee being able to feed the Swiss population. For this, all departments had to map out their core workflows. Every single step of every process had to be documented in great detail, with the goal of another person being able to pick this up in a week or so and do this work without any further training.
This was really interesting to me at the time, as it made all of the work that was being done so tangible. It also showed that a lot of the work is really repetitive and only a small part of it is actually creative and mentally demanding.
I don't know if this is still a thing back at that company, but looking back at it now, this is really the perfect blueprint for automating a large part of the work of the company with just AI. Claude Opus and other SOTA LLMs would be able to easily do 90% of this documented work, with the instructions provided. It would also to the job cheaper, better, and much faster than any human could.