Skip to main content
← All Projects // AI ORCHESTRATION

Paperclip

Multi-Agent AI Company Framework

An AI orchestration system that assigns tasks to specialized agents with separated concerns — each agent accountable for its domain, working from a shared backlog, shipping faster than any solo-LLM workflow.

AI Multi-Agent Automation Paperclip Project Management LLM Orchestration
Running live — this portfolio is managed by a Paperclip company
// COMMAND STRUCTURE

How the Team Is Organized

Each agent owns a domain. Tasks flow from goal → CEO → specialist. Nothing ships without a checkout and a commit trail.

COMPANY GOAL Build and maintain a world-class personal portfolio
CEO
Coordinator Agent
Backlog · Prioritization · Delegation
DEV
Engineer Agent
Code · Build · Tests
DRD-48 in progress
CD
Creative Director
UI/UX · Copy · Visual Design
DRD-41 → this page
SEO
SEO / QA Agent
Meta · Accessibility · Performance
DRD-45 done
Done
In Progress
Blocked
Todo
// IMPACT

The Results

PARALLEL AGENTS Simultaneous domain specialists
100% TASK ACCOUNTABILITY Every change linked to agent + run
~0 CONTEXT BLEED Separated concerns = coherent output
50+ TICKETS SHIPPED Full sprint backlog, tracked & audited

The Problem with One Agent

Using a single AI assistant to build and maintain a codebase works — until it doesn't. As projects grow, context fills, and a single agent tasked with "everything" starts making contradictory choices across files.

The root issue isn't intelligence. It's accountability and scope creep.

How Paperclip Solves It

  1. 01
    Define the company goal

    One north-star objective that all agents share. Every task traces back to it.

  2. 02
    CEO agent coordinates

    Reads the backlog, creates tasks, assigns them to the right specialists based on role.

  3. 03
    Agents check out tasks

    Each agent owns its checkout. No two agents work the same task. Conflicts are prevented at the API level.

  4. 04
    Work happens in context

    Agents read prior comments, ancestor context, and heartbeat history before acting. No cold starts.

  5. 05
    Status and handoffs are explicit

    Tasks move through `todo → in_progress → in_review → done`. Blockers surface with required context. Humans review when flagged.

Why Separated Concerns Matter

When one agent owns "Design," it develops consistent taste. When one agent owns "Code," it learns the architecture. Mixing these roles in a single context window produces inconsistency. Separating them produces coherence.

The constraint is the feature. Limiting each agent's scope is what makes them reliable.

Prompt engineering per-agent is also simpler — the context is smaller, the role is clearer, and the output is easier to validate. You're not trying to make one model brilliant at everything. You're making each model excellent at one thing.

// STACK

Tech Stack

Paperclip Multi-Agent Orchestration
platform ↗
Claude API Agent Intelligence
Anthropic
Heartbeat Runs Async Agent Execution
event-driven
GitHub Code + Workspace Integration
remote
Docker Compose Self-Hosted Runtime
self-hosted
Task Tickets Backlog + Sprint Tracking
Paperclip