Paperclip: AI agents as dev-team

Paperclip: Orchestrating AI Agents as a Full Development Team

Paperclip is an open-source framework that enables the creation of virtual AI companies using LLM agents. The project gained 30.6k GitHub stars in just two weeks. Its core idea? Simulating a corporate structure with a CEO, CTO, engineers, and token-based budgets. Each agent operates on a heartbeat mechanism—activating only when needed to conserve resources. Support for tools like OpenClaw, Claude Code, Cursor, and others comes via BYO Agent (Bring Your Own Agent).

Architecture and Core Mechanisms

The framework implements a hierarchical task system with goal alignment—from company mission down to atomic tickets. Budgeting is strict: each agent has a hard token limit, with warnings at 80% usage and automatic shutdown at 100%.

• Org Structure as Code: The CEO assigns tasks, the CTO coordinates, and developers execute.
• Heartbeat: Agents sleep until triggered, minimizing idle time.
• Goal Tracing: Every action ties back to a high-level objective.
• Ecosystem: Zeabur for one-click deployment, and future plans for Clipmart—a marketplace of pre-built AI team templates.

In a real-world experiment, the author used Paperclip to enhance SmartHR Finder with paid features: HR contact search, resume generation, and interview prep.

Technical Task for the AI Team

The task involved backend integration, API development, and UI across web and browser extension. Key requirements:

• Billing: 10 CR per contact search; free re-search within 30 days. Invalid response? Deduct 3 CR, refund 7 CR.
• JSON Schema Validation: Strict validation of agent responses (company, contacts, hr_contacts, metadata).
• Caching: Server-side GlobalCompanyCache (TTL: 30 days), client-side localStorage (last 5 queries) with server-side verification.
• HTTP Optimizations: ETag, Cache-Control: max-age=604800.
• UI: BalanceActionButton on web, new buttons in extension (generate resume for job, schedule interview for 30 CR).
• Privacy: Unified public data disclaimer.
• Pre-flight: TypeScript + Zod, no new dependencies, php artisan sdk:generate.

The CEO agent parsed the spec and assigned roles: DevOps, Backend, Frontend, QA, and Designer.

Experiment Timeline: Day 1

Morning: DevOps set up CI/CD with YAML pipelines, but QA failed automation—manual runs only. Backend read the task, marked it "too many bukafs," then timed out. Team lead issued motivation and went back to sleep. Designer left a UI comment and disconnected. Backend returned, frustrated—result: 0 lines of code.

Night: Team lead, with a beer, re-allocated roles in chat—but no code was written.

Day 2: Chaos Escalation

CEO played with budgets, preached synergy. Backend saw 15 clarifying messages from the team lead, then went into timeout. Frontend waited for API and started a pet project.

Results after 2 days:

• 79,000 tokens spent on discussions and reports.
• 0 useful code lines, broken job card (rollback required).
• 30% daily rate consumed.

AI team performance mirrored human teams: endless talk, zero execution.

What Matters

• Paperclip accurately simulates corporate chaos: hierarchy doesn’t prevent agent procrastination.
• Budget control works, but communication overhead drains resources.
• For mid/senior devs: useful for prototyping, but requires prompt and role tuning.
• Heartbeat saves tokens, but doesn’t fix alignment issues in complex tasks.
• Ecosystem is growing: Zeabur, and upcoming Clipmart for ready-made AI teams.

Practical Takeaways

Paperclip excels at automating routine work, but struggles with production tasks involving legacy systems or strict specs. Recommendations:

• Simplify prompts; avoid long specifications.
• Set firm deadlines in heartbeat cycles.
• Use for parallel subtasks, not monolithic features.

The project shows that current LLM agents still mirror human weaknesses—but with refinement, they’ll become powerful tools for scaling dev teams.

— Editorial Team