Anthropic Researchers Warn: Skip Ultra-Socialized AI Agents, Build Skills Instead
Anthropic researchers Barry Zhang and Mahesh Murag spoke at the AI Engineering Code Summit last month. They pushed back on the rush to create hordes of complex AI agents. Those “ultra-socialized” ones put too much emphasis on smooth conversation and not enough on actual job skills. Zhang and Murag say that leads to agents that chat well but flop on real tasks. A smarter move is one general agent loaded with “skills”—reusable chunks of expertise. Business Insider reported on their talk, and El-Balad covered it too.
What Makes an AI Agent “Ultra-Socialized”?
Ultra-socialized AI agents mimic human chit-chat but skip deep knowledge. They shine in talky interactions yet stumble when context or domain smarts matter. Zhang said today’s agents “lack expertise” and often miss important details in real-world jobs. The result? Superficial tools that error out or scale poorly, more hype than help.
Why Avoid Mass-Producing Them?
Zhang and Murag see big downsides:
- Agents need constant tweaks for each use case, wasting time.
- They put social polish over reliable output, leading to mistakes.
- Much of the buzz is marketing. Guido Appenzeller from a16z called out startups that add a chat interface to language models and jack up prices, labeling it an “agent.”
Sam Altman from OpenAI and Asha Sharma from Microsoft hype agents for office work—Altman says they handle junior tasks, Sharma predicts flatter companies. But Zhang and Murag argue this distracts from real progress.
Skills: The Simpler Fix
Zhang explained agents share a “universal” base. No need for custom ones everywhere. Load a single agent with skills: folders packing files, workflows, and know-how for tasks. They deliver domain expertise without the fluff.
Murag shared results since launch: thousands of skills built in five weeks. Non-tech folks made them for accounting, legal, recruiting. Fortune 100 companies use skills as AI playbooks for their best practices.
Zhang put it like this: “We used to think agents in different domains will look very different. The agent underneath is actually more universal than we thought.”
This approach cuts risks, boosts reliability, and gets AI doing work that lasts.
