Tech • AI • Robotics • Game
Anthropic has stopped creating agents: here’s what they’re doing instead
7/10TL;DR
Anthropic-style agent workflows are shifting away from one agent per task and toward a single assistant extended with reusable skills that package procedures, tools and scripts for repeated work.
KEY POINTS
From many agents to one generalist
A growing practice in AI operations is to stop creating separate agents for each business function and instead maintain one main assistant that can perform many jobs through add-on skills. The model is closer to installing software on one console than buying a new console for every game: the assistant stays the same, while capabilities are loaded when needed.
Why the multi-agent trend is inefficient
Building isolated agents for thumbnails, ads, copywriting or coding often leads to duplicated setup, repeated prompting and inconsistent results. The alternative is a single assistant that can invoke the right procedure for a task, reducing the need to restate instructions every day and avoiding fragile memory assumptions about past chats.
Claude Code over basic chat
The workflow described relies on Claude Code rather than a limited chat interface. The coding environment gives the assistant broader access to files, scripts and browser actions, making it practical to automate production tasks instead of only answering prompts in a minimal text window.
What a skill actually is
A skill is essentially a reusable procedure file, sometimes backed by scripts, that tells an assistant how to execute a specific task in a repeatable way. Unlike a permanent memory file that is read on every request, a skill is called only when relevant, which keeps the system lighter and more targeted.
A concrete use case: YouTube thumbnail generation
One example automates the remake of underperforming YouTube thumbnails. The assistant receives a video URL, searches for high-performing thumbnails on the same topic, generates multiple new options and prepares them for review. In one case, it produced 20 variants for selection, compressing work that previously could have taken months into a process launched in seconds.
Skills can chain tools together
The thumbnail workflow combines separate skills, including one for image generation with Codex and another for scraping competitive examples with TRG. That means a user does not need a dedicated “thumbnail agent”; the main assistant simply composes the necessary skills for the job.
Iteration comes before formalizing the skill
The recommended method is to first perform the task manually with the assistant in natural language, then refine it step by step. In the thumbnail case, the workflow evolved to include an interactive HTML selector, title suggestions, title auditing against web best practices and a cleaner interface before being saved as a formal reusable skill.
Interactive outputs improve review
Rather than dumping files into a folder, the assistant can build a lightweight interface where users click through generated thumbnails, compare options and pair each visual with headline choices. That turns the assistant into a working production tool, not just a text generator.
Economics matter
The setup also emphasizes cost control. Using Codex image generation within a $20 monthly subscription was presented as a way to avoid per-image charges that can run around 14 cents each, making high-volume experimentation more viable.
Skills are portable but require caution
Skills can be stored in dedicated directories and reused across projects or external assistants, including local and remote setups. But because a skill may include executable commands and scripts, importing skills from the internet carries security risk and requires review before use.
CONCLUSION
The larger shift is from prompt-by-prompt micromanagement to documented, reusable AI procedures. For businesses, the main advantage is not more agents, but one assistant that becomes more capable over time through a growing library of skills.
Ask a question