Paperclip promises a company of AI agents, but in testing it drowned in bureaucracy

Paperclip offers to manage not a single AI assistant but a whole "company" of agents with roles, budgets and an internal hierarchy. A Habr author tested the idea on a real product task and got a telling failure: over two days, a team of eight virtual "employees" spent tokens, generated noise and got almost nothing done.

How Paperclip works

Paperclip is an open-source project that, according to the author, collected 24,000 GitHub stars in its first 12 days and had reached 30,600 by the time the article was published. Its idea is simple and ambitious at once: stop treating AI as a set of separate chats and turn the interaction into a company model. Instead of one assistant, the user gets a CEO, a CTO, developers, a designer and other "employees" among whom tasks, budgets and responsibility can be divided.

"If OpenClaw is an employee, then Paperclip is a company."

• roles such as CEO, CTO, engineers and designers for different types of tasks
• token limits, with agents switched off when the budget runs out
• a heartbeat mechanism that wakes agents on a schedule instead of keeping them constantly active
• a link between every task and the overall goal, so the agent understands the context of its work
• support for different tools, from Claude Code and Codex to Bash and HTTP agents

On paper this looks very powerful. Paperclip tries to solve a real pain for developers who have dozens of agent sessions open at once, lose context and watch token costs spiral out of control. An ecosystem is already forming around the project: cloud deployments, an active community and the idea of a marketplace of ready-made "AI companies" for typical scenarios, from marketing to development. That explains why interest in the project grew so fast.

Why everything stalled

For the test, the author gave Paperclip a perfectly ordinary task: finish an existing product, add a couple of endpoints, integrate an external API and bring the UI in line with an already drawn design. The system then behaved all too much like people.

The CEO handed out tasks and vanished to look at metrics. DevOps got carried away with automation and generated a pile of CI/CD configs. QA refused to work with that format. Backend saw the size of the specification and irritably backed off. The designer limited himself to a remark along the lines of "make the buttons like in other places." The team lead woke up from time to time, issued strategic guidance and went away again.

The second day was no better. Backend came back, saw the new clarifications and a long thread, then went into a deep timeout. Frontend sat waiting for an API it never received.

The outcome was almost satirical: in two days, eight agents spent 79,000 tokens, wrote not a single useful line of code, broke the job listing card and drove the author to roll back the changes. After that, a lone contractor in the form of Cursor closed the same task in about an hour and a half.

Where the real value is

The author makes an important caveat: the problem is not that Paperclip is useless. On the contrary, it hit a very precise spot in the market. When a developer has Claude Code, Cursor, Codex and other agents running at once, chaos sets in quickly: it is unclear who is doing what, where the discussion history is, what context has been lost and how much money has already been burned. Paperclip offers a management layer on top of this zoo, with tickets, auditing, state recovery and budget control.

But the experiment also shows the main risk. Once interaction is built on the model of an ordinary company, bureaucracy, coordination losses and dependence on the quality of each individual agent arrive along with the useful division of labor. Paperclip does not make models smarter and does not fix bad decisions; it only organizes them into a structure. So the most sensible role for a human in such a system today is neither micromanager nor "employee next to the bots" but a board of directors: set the goal, allocate the budget, watch the metrics and step in only where autonomy genuinely breaks down.

What this means

Paperclip has correctly identified a coming problem for the market: managing several agents is already harder than working with one. But the article clearly shows the limit of current capabilities: coordination tools are needed now, but copying a human company with its hierarchy and rituals for autonomous agents is still premature.