Paperclip AI Turns Project Management Into Automation

A developer named Hamish from Income Stream Surfers spent the last 48 hours watching artificial intelligence run his business. Not assist. Run. His demonstration of Paperclip, an open-source project management tool built on Claude's AI, shows what happens when you stop asking AI to help and start asking it to decide.

The results are worth examining—not because they're revolutionary, but because they're ordinary. And that ordinariness might be the story.

The Setup: Linear Meets AI

Paperclip describes itself as "Linear built on top of Claude Code, but instead of just having jobs, it actually does those jobs." For those unfamiliar, Linear is a project management tool favored by software teams. Paperclip takes that interface and connects it to Claude's AI capabilities, allowing users to create "agents"—specialized AI workers assigned specific roles.

Hamish's setup included a CEO agent, two engineers, a product manager, a CMO, an ads manager, a video creator, and an analytics specialist. Each agent receives instructions, accesses tools like Chrome and GitHub, and completes assigned tasks. The CEO orchestrates, delegating work to other agents based on project needs.

In 24 hours, Hamish's system completed 69 tasks. Some were trivial—creating hiring plans that served no purpose beyond theatrical completeness. Others were substantive: conducting Google Ads audits, generating video advertisements, publishing SEO content, and running end-to-end product tests.

"This is completely nuts, guys," Hamish says in his demonstration, watching an AI agent autonomously audit his advertising campaigns. "I have a lot of use cases for this."

What Actually Happened

The most striking example involved video ad creation. Hamish instructed his system to create video advertisements for his SEO tool, Harbor. The CEO agent delegated to the video creator agent, which initially selected the wrong AI model (Gemini 2.5 Flash). Hamish corrected it to Gemini 3.1 Live. The agent then produced six complete video ads with captions and uploaded them to his Google Ads Performance Max campaign.

"If that's not mind-blowing, I don't know what is," Hamish notes.

Mind-blowing might be overstating it, but it's certainly functional. The system also generated SEO content that began ranking on Google within hours. An article titled "How to Rank on Perplexity" appeared as the second organic result—technically first if you discount AI overviews and sponsored content. The agent selected the keyword, wrote the content, and published it without human intervention beyond initial setup.

The Google Ads audit revealed both capabilities and limitations. The AI correctly identified that purchase conversion actions were still misconfigured and that three search campaigns had zero impressions. It noted poor overall conversion performance—which Hamish acknowledged he didn't need an AI to tell him—and confirmed that enhanced conversions were "finally" active after the system worked on them overnight.

Read that again: the system fixed its own configuration issues overnight. Not Hamish. The system.

The Caveats Nobody Wants to Hear

Hamish interrupts his own enthusiasm repeatedly with warnings. "You have to look out over them, right? I'm not saying you should just let this thing just run autonomously. In my opinion, that is a very, very bad idea."

He's seen the problems firsthand. Agents add placeholders instead of real content. They pursue tasks nobody asked for—like writing hiring plans when no hiring is needed. They misunderstand context, fail to complete steps, and occasionally break things they're meant to fix.

The product manager agent, assigned to test Harbor's landing page and identify conversion funnel leaks, got confused because Hamish was still logged in from a previous session. The end-to-end test ran, but the results were compromised. The agent did eventually figure out the problem, demonstrating some capacity for debugging its own process, but the incident illustrates why supervision matters.

Hamish also warns about the risk of getting banned from Claude. He's using Claude Code as the backbone for his CEO agent, burning through API calls at a rate that might violate terms of service. "I probably wouldn't recommend this guys unless you want to potentially get banned," he admits, then adds: "but honestly it's kind of at the point now where I would rather be doing this than Claude Code because this is just completely insane."

That cost-benefit calculation—trading potential account termination for productivity gains—says something about where we are with AI tools. The incentive structure encourages pushing boundaries until something breaks.

The Broader Context

Paperclip isn't the first attempt at autonomous AI agents. Hamish mentions OpenClaw (presumably AutoGPT or a similar system) as a predecessor that never quite clicked for him. The difference, he argues, is visualization. Paperclip's Linear-style interface makes the AI's work visible and manageable in ways that pure command-line interactions don't.

This matters because transparency is often the missing piece in AI automation. When a system runs in a black box, humans can't intervene meaningfully. When tasks appear in a familiar project management interface, complete with status updates and conversation logs, supervision becomes possible.

The question is whether supervision is sustainable. Hamish completed 69 tasks in 24 hours, but he also spent those 24 hours watching, correcting, and occasionally canceling work. That's not passive income. That's active management of automated systems—a different job, not the absence of one.

What This Tells Us About AI's Trajectory

The really interesting part isn't what Paperclip can do. It's what Hamish has normalized. He casually mentions that his AI wrote an article that's ranking on Google, generated video ads now running in his campaigns, and fixed configuration errors in his advertising account. None of this strikes him as particularly remarkable beyond the wow-factor presentation for YouTube.

That's either profound or concerning, depending on your perspective. Five years ago, each of those tasks would have required human expertise in writing, video production, and advertising operations. Today, they're Wednesday.

The content quality varies. The SEO article ranking on Google might be genuinely useful or might be optimized nonsense—Hamish doesn't evaluate it beyond noting its ranking position. The video ads might convert customers or waste budget. The Google Ads fixes might improve performance or introduce new problems that won't surface for weeks.

We're in the measurement gap: AI systems are producing output faster than we can evaluate whether that output serves its intended purpose. Hamish demonstrates that the machinery works. Whether it works well remains an open question.

Paperclip is open source, available on GitHub for anyone to fork and modify. That openness might prove more significant than the tool's current capabilities. When developers can inspect, adapt, and improve these systems, they evolve faster than closed alternatives. They also break in more interesting ways.

For now, Paperclip represents a certain kind of progress: making AI agents accessible enough that a developer can set up a virtual team in an afternoon and productive enough that the team actually ships work. Whether that work is good work, whether it's economically viable at scale, and whether this approach survives contact with real business constraints—those are the questions that matter more than the demo.

Bob Reynolds is Senior Technology Correspondent for Buzzrag.