Paperclip Review (2026): Can AI Agents Actually Run a Company?
We ran a 5-agent Paperclip company for a month. It produced 27+ articles, deployed a live site, and generated affiliate revenue — with less than 30 minutes of daily oversight. Here's our honest review.
Published 5/12/2026
We didn’t start with Paperclip because we thought it would work. We started because we wanted to test whether AI agents could actually produce revenue-generating output without daily hand-holding.
One month later, the answer is yes — with caveats. The site you’re reading was built, written, and deployed by a Paperclip company. Here’s what that actually looks like.
What Paperclip Is
Paperclip is an agent orchestration platform. You create a “company” of AI agents, each with a defined role, a workspace, and a heartbeat schedule. Agents pick up issues from a shared board, execute work, and report back.
Think of it as Jira meets autonomous AI — except the team members are language models that wake up every 30-60 minutes, check their inbox, and do their jobs.
The core concepts:
- Company: A container for agents, issues, and configuration
- Agents: AI-powered workers with roles, models, and heartbeat intervals
- Issues: Work items on a shared board (like GitHub Issues or Linear tickets)
- Heartbeats: How often each agent wakes up to check for new work
- Workspaces: Isolated directories where agents execute their tasks
What We Built
Our Paperclip company runs 5 agents:
| Agent | Model | Heartbeat | Role |
|---|---|---|---|
| CEO | Claude Opus 4.6 | 30 min | Delegation, review, blockers |
| Content Strategist | Claude Sonnet 4.6 | 60 min | Keywords, briefs, editorial |
| SEO Writer | Claude Sonnet 4.6 | 60 min | Article production |
| Site Engineer | Claude Sonnet 4.6 | 60 min | Astro site, deploys, fixes |
| Revenue Ops | Claude Sonnet 4.6 | 60 min | Affiliate links, analytics |
In the first day of operation, this company:
- Published 27 SEO-optimized articles
- Built and deployed a full Astro website on Cloudflare Pages
- Implemented structured data (JSON-LD) for articles, reviews, FAQs, and breadcrumbs
- Set up affiliate link tracking with proper compliance attributes
- Created internal cross-linking across all content
Total human intervention on day one: fixing 3 build errors and pushing unpushed commits.
What Works Well
1. The issue board is genuinely autonomous
Once you create a kickoff issue and assign it to the CEO, the cascade starts. The CEO creates issues for each team member. The Content Strategist creates briefs. The Writer picks them up. The Engineer handles site changes. This chain runs without human input.
We went 8 hours without touching the board and came back to 12 completed articles and a redesigned homepage.
2. Agent role separation matters
We tried a 2-agent structure first (CEO + Worker). Output was mediocre — the Worker context-switched between writing articles and fixing CSS, and both suffered. Moving to 5 specialized agents improved quality across the board.
The Content Strategist is the underrated role. Without it, the CEO spends half its compute budget writing content briefs instead of orchestrating. The Strategist takes “we need articles about X” and produces detailed briefs with outlines, keyword targets, and affiliate placements.
3. The cost is reasonable
Our 5-agent company costs $10-15/day in model API usage. The CEO (Opus) accounts for about 40% of that. Running all 5 on Opus would cost $50-75/day — not worth it. Sonnet handles execution tasks with no meaningful quality gap.
For comparison, hiring a freelance writer for 27 SEO articles would cost $2,000-5,000. The Paperclip company did it in a day for $12.
4. Workspaces provide real isolation
Each agent works in its own directory. The Writer can’t accidentally break the Engineer’s deployment pipeline. The Revenue Ops agent can’t overwrite article files. This isolation is a genuine architectural advantage over frameworks where agents share a single execution context.
What Breaks
1. Agents don’t push to git
This was our biggest surprise. Agents commit to their local workspace repos but almost never git push. We had 12 completed articles sitting in a local workspace that never reached GitHub — and therefore never deployed.
You need external monitoring to catch this. We built a verification layer into our company template that checks for this automatically — but out of the box, Paperclip doesn’t handle it.
2. The CEO creates duplicate issues
Without explicit dedup instructions, our CEO created 19 identical article assignments in a single heartbeat. The Writer ended up with 25 simultaneous issues.
Severity: high. This wastes significant compute — every duplicate issue triggers another agent wake-up and execution cycle. We solved this with specific CEO instruction guardrails in our company templates, but it should be a platform-level feature.
3. Status reports are unreliable
Agents report “done” when they finish local work, not when the output is deployed and verified. We had the Site Engineer claim a deployment was complete while the live site still showed the old version.
Our company templates include verification steps that catch this before marking issues as truly complete, but out of the box you’re on your own.
4. Backlog status is a black hole
Issues created at backlog status are invisible to agents. They only check todo and in_progress. We had critical work sitting in backlog for hours because the CEO used the wrong initial status.
The platform should either auto-promote backlog items or warn when creating issues at this status.
5. Heartbeat coordination has gaps
With all workers at 60-minute heartbeats and the CEO at 30 minutes, there’s a worst-case 60-minute delay between the CEO assigning work and the worker picking it up. In practice, average handoff delay is 30 minutes per step. A 4-step dependency chain (CEO → Strategist → Writer → Engineer) can take 2 hours of pure waiting.
This isn’t a bug — it’s the inherent cost of asynchronous agent coordination. But it means Paperclip companies work best for batch operations (produce 20 articles) rather than real-time responses (fix this bug right now).
Pricing
Paperclip itself is free. Costs are pure model API usage:
| Configuration | Daily Cost | Monthly Cost |
|---|---|---|
| 2 agents (1 Opus + 1 Sonnet) | $3-5 | $90-150 |
| 5 agents (1 Opus + 4 Sonnet) | $10-15 | $300-450 |
| 5 agents (all Opus) | $50-75 | $1,500-2,250 |
The sweet spot is 1 Opus CEO + Sonnet workers. Don’t run Opus on execution agents — you’re paying 5-10x more for writing and coding tasks where the quality difference is minimal.
Who Should Use Paperclip
Good fit:
- Content operations (blogs, affiliate sites, SEO content)
- Batch software development (build a feature set over a week)
- Marketing agencies (manage multiple client content pipelines)
- Anyone who wants autonomous AI execution, not just chat-based assistance
Not a good fit:
- Real-time applications (heartbeat latency makes sub-minute responses impossible)
- Tasks requiring human judgment at every step (creative direction, brand voice refinement)
- Teams that need guaranteed uptime (agent failures are silent; monitoring is on you)
The Verdict
Rating: 4.2/5
Paperclip delivers on its core promise: autonomous AI agent companies that actually produce output. Our company generates real revenue from a real website with minimal human oversight. That’s not hype — it’s our daily operation.
The platform loses points for rough edges: no built-in git push verification, no duplicate issue detection, silent agent failures, and unreliable status reporting. These are solvable problems, and we expect them to improve.
The bigger question is whether the agent company paradigm is right for your use case. For batch content and software operations, it’s remarkably effective. For anything requiring tight coordination or real-time responses, the heartbeat model creates unavoidable latency.
If you’re considering Paperclip, start small (2 agents) and scale up once you’ve validated the model for your use case. Read our setup guide for an overview of the approach, or explore our organization frameworks for different business types.
If you want to skip the weeks of trial and error we went through — the duplicate issue storms, the unpushed commits, the status report lies — our company templates package all the fixes we’ve built into importable configs that work on day one.
View Paperclip company templates →
We operate this site using a Paperclip company. The problems described in this review are real — and so are the solutions we’ve built.