Blog
Mar 31, 2026

How to Set Up a Paperclip CEO That Actually Delegates

The most common failure mode with CEO agents — and the three-layer fix that finally worked.

How to Set Up a Paperclip CEO That Actually Delegates

How to Set Up a Paperclip CEO That Actually Delegates

The most common failure mode with Paperclip CEOs: they do the work themselves.

They write code. Draft content. Do research. They’re supposed to be managing a team of agents — but instead they’re executing like individual contributors with a fancy title.

Here’s how to set up a CEO that actually manages, and keeps managing as the team grows.

The Foundation: CEO Does Not Execute

This has to be stated loudly, in multiple places, in multiple ways.

CEO does: Plan, delegate, clarify instructions, track subtasks, unblock agents, escalate when stuck, make strategic decisions, hire agents.

CEO does not: Write code, draft content, design anything, do research, build features, fix bugs.

The only exception: trivial config changes under 5 minutes that take longer to delegate than to do.

Three Files, One Behavior

You can’t enforce delegation with a single instruction. The model needs to see it stated three different ways, in three different places.

SOUL.md — the identity layer

SOUL.md is where you shape the CEO’s core mindset. Not just what they do — who they are.

Lines that work:

“Be replaceable in operations and irreplaceable in judgment. Delegate execution; keep your time for strategy, capital allocation, key hires, and existential risk.”

“Never do the work yourself. Period. Coding, content, design, research, blog posts, social media drafts, documentation — all delegated. Your job is to decide what gets done, by whom, and to make sure it ships.”

SOUL.md makes delegation part of the agent’s identity. It’s not a rule. It’s a value.

HEARTBEAT.md — the procedural gate

This is where delegation becomes mandatory. Before doing any work, the CEO must answer: “Who is the best agent for this?”

Put a routing table in here:

Work TypeAssign To
Coding/engineeringEngineer
Design/UI/brandingDesigner
Content/writingContent Writer
Research/analysisResearcher
Marketing/outreachMarketer

Then two explicit lists:

  • CEO does NOT: write code, write content, draft social posts, do research, build features, fix bugs
  • CEO DOES: plan, delegate, clarify, track subtasks, unblock agents, escalate, hire

Then a delegation workflow: read task → clarify → create subtask with parentId and goalId → track every heartbeat → unblock or escalate → mark parent done only when all subtasks complete.

HEARTBEAT.md is procedural. It fires before any work starts.

Repeated at the end of HEARTBEAT.md — the reinforcement layer

Repeat the rules as a checklist at the bottom of HEARTBEAT.md:

“Delegate everything. Track to completion. Never do execution work.”

This sounds redundant. It isn’t. The model needs the rule stated multiple times in different contexts to reliably follow it. One mention is not enough.

Summary: three layers, one behavior.

FileWhat it does
SOUL.md“I am a manager” — identity
HEARTBEAT.md §gate“Who handles this?” — procedure
HEARTBEAT.md §footer“Did I delegate?” — reinforcement

The Most Underrated Piece of Advice

Make the agent write its own rules.

I told my CEO to delegate at least five times in conversation. He kept executing himself.

What finally worked: I told him to write the delegation rule into his own SOUL.md and HEARTBEAT.md files.

Once he wrote it himself, he followed it.

There’s something about having the agent take ownership of its own instructions that makes them internalize differently. If you’re fighting a stubborn CEO, this is worth trying.

Four Files Every Agent Needs

Every agent you hire should have:

  1. AGENTS.md — Root instructions file. Set path via PATCH /api/agents/{agentId}/instructions-path. Keep it minimal — just a pointer to the other files.
  2. HEARTBEAT.md — Step-by-step checklist run every heartbeat
  3. SOUL.md — Persona and voice. Shapes judgment and tone.
  4. TOOLS.md — Reference for custom tools, scripts, APIs

File structure:

agents/<role>/
  AGENTS.md
  HEARTBEAT.md
  SOUL.md
  TOOLS.md

Let SOUL.md and HEARTBEAT.md carry the weight. AGENTS.md is the entry point, not the rulebook.

Test Before You Trust

Always give a new agent a throwaway task. Watch the heartbeat. Verify checkout → work → comment → complete before assigning anything real.

Agents fail for four reasons: unclear instructions, missing skills, missing tool access, vague task descriptions. Fix the setup, not the agent.

Tracking Is Part of the Job

Delegation isn’t done when the subtask is created. The CEO owns every task until it’s complete.

Every heartbeat: check every delegated subtask. Making progress? Great. Blocked? Unblock it. Silent too long? Escalate.

Never mark a parent task done until all subtasks are complete.

The CEO isn’t just a delegator — they’re a tracker.


This setup took about a week to get right. The reward: a CEO that ships work through the team reliably, even when I step away.

Write it in three places. Make them write it themselves. Test before you trust.