Less agents, more skills

Edition 144 | December 15, 2025

Throw in a slick landing page, and call a prompt an “agent”.

Congratulations, you’re now 90% of today’s “AI agencies”.

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• Why agent skills > building more agents is becoming the dominant model

• Google’s new Deep Research agent for autonomous web research

• Cohere Rerank 4 and the quiet RAG accuracy boost

• AI agents outperformed human hackers in a real security test

• New guides and courses for building agents

…and more

🔍 SPOTLIGHT

The industry has been racing to build specialized agents for every use case, but Anthropic researchers say that's the wrong approach. At the AI Engineering Code Summit, Barry Zhang and Mahesh Murag argued teams should build Skills, not agents: folders of instructions, scripts (tools), and documentation that any agent can load dynamically.

Originally, Anthropic expected they'd need separate agents for each domain: one for finance, one for legal, one for code. Instead, the insight came from building Claude Code. They discovered "the agent underneath is more universal than we thought." Claude Code turned out to be a general-purpose agent that could do financial analysis, scientific research, or document editing, as long as it had the right context loaded. The bottleneck is expertise, not intelligence.

The concept is simple but genius. If you think of any human having a base level of intelligence, the only thing separating a person doing one job from another is having the right "skills": instructions, tools, and some extra context. Anthropic’s Skills package exactly that: the step-by-step procedures, scripts, and domain knowledge that turn a generalist AI agent into a specialist.

So what is a Skill? A Skill is just a folder containing a SKILL.md file with literal instructions, plus any scripts (tools), templates, or assets the agent needs. They're progressively disclosed: at runtime, the agent only sees metadata until it decides to use a skill, protecting the context window. You can equip an agent with hundreds of skills without blowing up token costs. When Claude needs to create a PowerPoint, it loads the presentation skill. When it needs to create financial reports, it loads that skill instead.

And the concept is already in use. 

Five weeks after launch, thousands of Skills have been created. Fortune 100 companies are treating them as internal playbooks for AI, teaching agents about organizational best practices and bespoke internal software.

What's surprised Anthropic the most is who's building them. They expected developers to dominate. Instead, non-technical teams in finance, recruiting, accounting, and legal are creating skills: people who've never written code but can describe their workflows in markdown format.

Why this matters: The "blank slate" problem, where every session starts from zero, has been one of the biggest barriers to reliable agents. Skills make memory tangible and transferable. Claude can write down what it learns in a format future versions of itself can use. The goal, as Zhang put it, is that "Claude on day 30 of working with you is a lot better than Claude on day one."

What do you know how to do that you could teach an agent?

Keep learning and building!

—AP