🧠 AI Companies With No Humans? Paperclip, Claude Marketplace, and Karpathy’s Autoresearch

 🔥🔥🔥 Three Highly Curated AI Stories

1. Paperclip: The open-source OS for AI-run companies

A new project called Paperclip just dropped on GitHub, and the idea behind it is wild. It’s essentially a company operating system designed for AI agents.

Instead of running a bunch of disconnected agents across different tabs and tools, Paperclip lets you organize them into a single structure. Agents can have roles, reporting lines, budgets, and goals, just like employees inside a company.

The system includes org charts for agents, monthly budgets to control costs, ticket systems with audit logs, and tool-call tracing to see what each agent is doing. Agents can also run continuously on scheduled “heartbeats,” while you monitor everything from a dashboard.

In simple terms: if something like OpenClaw behaves like an employee, Paperclip acts like the company layer that manages all of them.

The project is already picking up traction with 1.4K stars and an MIT open-source license, which means anyone can experiment with it.

2. Claude Marketplace arrives for enterprises

Anthropic just announced the Claude Marketplace, a new platform designed to simplify how companies discover and procure AI tools built on Claude.

The idea is straightforward: instead of every enterprise building custom AI integrations from scratch, they’ll be able to browse and adopt pre-built tools and capabilities inside a centralized ecosystem.

For large organizations, procurement is often the biggest barrier to adopting new software. A marketplace approach could make it easier for teams to experiment with AI workflows without navigating long internal approval processes.

Right now the Claude Marketplace is in limited preview, so access is restricted while Anthropic tests how companies interact with the ecosystem.

3. Karpathy’s “Autoresearch” experiments with self-improving AI research

Andrej Karpathy shared a small but fascinating repo called Autoresearch, which explores what happens when AI agents start iterating on their own research loops.

The project packages a minimal LLM training setup into a single-GPU system with roughly 630 lines of code. From there, the human modifies the prompt instructions while the AI agent continuously updates the training code.

Each training run takes about five minutes, and the agent keeps experimenting with new configurations. It tweaks hyperparameters, architecture settings, and optimization strategies in search of lower validation loss.

Over time, the agent accumulates commits to the training script as it discovers better setups. The interesting part is the framing: instead of manually running experiments, you could imagine agents pushing research progress forward in an autonomous loop.

🔥🔥 Two AI Tools Worth Knowing

1.📘 Anthropic’s official guide to building Claude skills

Anthropic released a free guide explaining how to build “skills” for Claude, which are structured capabilities that allow the model to perform specialized tasks.

The guide walks through how to design skills, structure prompts, and integrate them into workflows. If you’re building AI tools, internal assistants, or automation pipelines, this document gives a clear look at how Anthropic thinks about extending Claude’s capabilities.

It’s one of the more practical resources if you want to move from “prompting” into actual AI product development.

2. 🧑‍💻 Agency Agents: spin up an AI agency with AI employees

A GitHub repo called Agency Agents is getting a lot of attention because of its framing.

Instead of building one giant AI agent that tries to do everything, the system organizes agents into departments like a company.

There are agents for engineering, design, marketing, product, testing, customer support, and even project management. Each agent specializes in a role, and they coordinate with each other to move projects forward.

The repo already has 10K+ stars in under a week, which suggests a lot of people are curious about the “AI company” model.

If you’ve been experimenting with multi-agent workflows, this is an interesting architecture to explore.

🔥 Things You Didn’t Know You Can Do With AI

Use Claude to find jobs and apply automatically

Claude now has browser-control capabilities through extensions, which means it can actually perform tasks inside your browser instead of just generating text.

One example making the rounds: using Claude to search job listings and apply automatically.

Here’s how it works.

1. Install the Claude Chrome extension from the Chrome Web Store.
2. Upload your resume to Claude.
3. Ask Claude to find high-paying jobs that match your experience.
4. Let Claude browse job sites, filter listings, and fill out applications automatically.

Because the model can control browser actions, it can handle steps like opening listings, filling forms, and submitting applications.

Watch a quick demo here.

Quick warning: If you automate too aggressively, platforms may flag your activity. Use it carefully.