Hierarchical AI Agent Delegation System
Run a 24/7 autonomous AI firm by issuing objectives to one CEO agent who manages the rest
The Hierarchical AI Agent Delegation System organizes an autonomous AI workforce around a single human touchpoint: a CEO agent. The human acts as a board member, issuing high-level objectives to the CEO, who interprets and delegates specific tasks to specialized sub-agents such as Research, Backtest, Risk Management, Execution, and Cost Optimizer. Each agent holds defined responsibilities and persistent memory, so the firm builds on prior outputs rather than restarting each session. The system is bootstrapped via a one-shot intake prompt in Claude Code backed by the open-source Paperclip framework. API costs are controlled by running tasks sequentially. Once strategies are validated in paper trading mode, a live exchange connection activates real execution—creating a genuinely autonomous operation the human need only supervise at the CEO layer.
- One interface rule: always communicate objectives through the CEO agent only, never bypass to sub-agents
- Role specialization: each agent has a narrow, explicitly documented set of responsibilities
- Memory persistence: agents log completed work so the firm learns and compounds over time rather than repeating tasks
- Cost discipline: run tasks sequentially or in small batches to prevent runaway parallel API spend
- Progressive validation: operate in paper trading mode first; connect live capital only after backtesting confirms the strategy
- Hierarchical accountability: results always flow back up through the CEO, who reports to the human board member
- Complete the intake interviewAnswer five structured questions in Claude Code: firm name, primary goal (grow capital, generate income, or test a strategy), strategy preference (existing or built from scratch), desired org chart (preset or custom), and risk tolerance. Speak or type in natural language—the system parses intent, not rigid syntax.Pro tipUse voice-to-text transcription to answer intake questions conversationally; the system handles imprecise language and will ask clarifying follow-ups if constraints conflict.WarningContradictory constraints—like 'never lose money' combined with 'moderate risk'—will be flagged by the system but not blocked, so resolve them deliberately or you will get a misconfigured risk agent.
- Select or design the agent org chartChoose a preset configuration (e.g., the six-agent setup: CEO, Research, Backtest, Risk Management, Execution, Cost Optimizer) or describe a custom hierarchy. Each agent gets a distinct role document outlining its responsibilities, which can be edited inside Paperclip at any time.Pro tipStart with the preset six-agent org chart for your first deployment; add custom agents only after you understand how task delegation flows between the preset roles.
- Run the one-shot setup prompt in Claude CodePaste the platform-specific one-shot prompt from the GitHub repository into Claude Code and click Always Allow on permission prompts. The script automatically checks for Node.js and Git, installs missing dependencies, creates the firm directory structure, and begins hiring agents via the Anthropic API.Pro tipEnable 'bypass permissions' in Claude Code settings before running to eliminate repeated permission dialogs during the multi-step install process.WarningDo not interrupt the setup mid-run; if node modules or Git are missing, the script installs them automatically—canceling partway through can leave dependencies in a broken state.
- Provide your Anthropic API key and connect data-source MCPsWhen prompted, paste your Anthropic API key from console.anthropic.com; the script stores it in the correct agent configuration folders automatically. Then connect at least one MCP (e.g., TradingView) so agents can pull live market data, run Pine Script backtests, and access research sources.Pro tipCreate a dedicated API key specifically for this agent firm so you can monitor spend and revoke access independently without disrupting other Claude Code projects.WarningMonitor your Anthropic API usage dashboard closely during the first run; running multiple agents simultaneously without sequential controls can consume $40 or more per minute.
- Verify agent activation in PaperclipOpen the Paperclip dashboard and confirm the expected number of agents appear as active on the left panel. Click into each agent's Instructions tab to review the role document the setup generated, and edit any responsibilities that do not match your intent before issuing the first task.Pro tipUse the Org view in Paperclip to visualize the delegation chain and confirm all sub-agents correctly point back to the CEO node before proceeding.
- Issue the first task to the CEO agent onlyIn Paperclip, click the CEO agent, select Assign Task, give the task a name, paste your strategic objective into the description field, and create the issue. The CEO agent will interpret your objective, break it into sub-tasks, and delegate them to the appropriate specialist agents without further input from you.Pro tipFrame your CEO task as a strategic objective, not a technical instruction—e.g., 'Research zero-drawdown yield strategies and propose three options with historical data' rather than 'call the research agent and search for X.'WarningNever assign tasks directly to sub-agents; bypassing the CEO breaks the delegation chain and prevents the firm from building coherent context across tasks.
- Monitor CEO inbox and iterate via commentsCheck the CEO agent's recent issues list to see task statuses. When the CEO marks a task complete and posts results to your inbox, leave a comment directing the next objective—for example, 'Now backtest the top two strategies over the past 24 months.' The CEO will re-delegate accordingly.Pro tipSet a calendar reminder to check the CEO inbox every few hours during initial setup; once the delegation rhythm is established, checks can become daily.
- Graduate from paper trading to live executionAfter the Backtest agent has validated a strategy with satisfactory historical results and the Research agent has confirmed current market conditions, disable paper trading mode in your exchange settings and connect your exchange API key to Claude. The Execution agent will then be authorized to place real trades.Pro tipRequire at least two full backtest cycles with documented results before disabling paper trading—one broad research pass and one focused optimization pass.WarningConnecting a live exchange API before paper trading validation is complete exposes real capital to an untested strategy; treat paper trading graduation as a formal checkpoint, not an optional step.
Lewis Jackson issued a single CEO task requesting passive income strategies with minimal drawdown. The CEO delegated to the Research agent, which autonomously searched the web for T-bill yields, money market fund rates, and Aave compound lending data for 2026. Upon completing its brief, the Research agent reported back to the CEO, which then commissioned the Backtest agent to validate a candidate strategy by uploading a Pine Script to TradingView and running historical tests—all without human involvement beyond the initial task.
The creator scaled the same hierarchical architecture to 32 agents organized into five departments covering research, backtesting, risk management, trade execution, and cost optimization. The CEO agent coordinates all five departments, memory persistence prevents redundant work, and the human operator interacts solely through the CEO layer or directly in Claude Code—never touching individual agents.
A content creator applies the same six-agent template—replacing Execution with a Publishing agent and Backtest with an Analytics agent—to run an autonomous content research and publication pipeline. The CEO receives a weekly topic brief, delegates research to the Research agent, routes drafts to a Review agent, and publishes based on Analytics agent performance benchmarks.
Extracted from Lewis Jackson, who built and documented a 32-agent autonomous trading firm using Claude Code and the open-source Paperclip multi-agent framework, then created a simplified six-agent entry-level version with a single copy-paste setup prompt to make the architecture accessible to non-developers.