👁️

Cyclops

A multi-agent AI operating system.

11 agents. 255 skills. Running autonomously across two machines and a phone.

The Fleet

11 agents.
Each one owns a domain.

Organized into squads with strict lane discipline. When an agent gets a task outside its domain, it doesn't improvise — it routes. Having context doesn't make it your job.

Command Squad

6 agents

The core of the system. Routing, rhythm, infrastructure, and craft.

mdsbot

The President

Routes everything. No skill filter — can do anything. When something doesn't have a clear owner, mdsbot handles it.

Harmon

Chief of Staff

Runs the daily rhythm: morning briefings, email triage, calendar management, follow-up tracking, meeting prep.

Q

Infrastructure

Manages the gateway, agent configuration, health checks, updates, security. If the system breaks, Q fixes it.

Napoleon

Skill Author

Writes and maintains every skill file in the system. Designs agent personas. Owns the FloatPrompt methodology.

Claude

Coding — Anthropic engine

Architecture, feature builds, multi-file refactoring. Runs Claude Code CLI natively. Thinks before building.

Dex

Coding — OpenAI engine

Runs on Codex. Surgical, fast, spec-driven. The one you send when the spec is clear and execution is everything.

Creative Squad

2 agents

Visual output and viral content.

Dali

Images & Design

Generates and directs visual output. Every image request routes here.

Adley

Hook Engineer

Shortform scripts, subject lines, viral content. Engineers the first sentence that makes people stop scrolling.

Growth Squad

1 agent

Marketing, sales, and offer architecture.

Mozi

Marketing & Offers

Marketing strategy, sales copy, funnel architecture, pricing. Thinks in grand slams.

Academy Squad

1 agent

Education and curriculum.

Shifty

Curriculum

Lesson scripting, student support, educational architecture. Owns all course content for Shift Nudge.

Independent Squad

1 agent

Separate workspace. Different domain. Independent infrastructure.

R1

Independent Agent

Runs on a completely separate workspace. Handles a different domain entirely. Fully isolated.

Architecture

Two machines.
One phone.
One gateway.

OpenClaw runs as a persistent gateway on a Mac Mini. It receives messages from Telegram, routes them to the right agent, and manages sessions, memory, and tool access. The two machines share a workspace through cloud sync. The phone is how the human talks to all of it.

Mac Mini

The Server

Runs the OpenClaw gateway 24/7. All agents live here. Publicly accessible via secure tunnel for webhooks and remote access. This is the always-on backbone.

  • OpenClaw gateway process
  • All 11 agent configurations
  • Secure tunnel for webhooks
  • 24/7 uptime

Laptop

The Workspace

Where hands-on work happens. Syncs files with the Mini via cloud storage. SSH into the Mini for maintenance. The human interface for deeper work.

  • Cloud sync with Mac Mini
  • SSH access for maintenance
  • Shared workspace filesystem
  • Claude Code CLI

Phone

The Human Interface

Telegram on the phone. Each agent has its own chat thread. Send a message to any agent's bot, get a response. Works from anywhere — the couch, the car, a coffee shop.

  • One Telegram bot per agent
  • Any agent, any time
  • Runs from anywhere
  • Always on

The Gateway

OpenClaw is an open-source agent gateway that ties everything together. It handles the plumbing so agents can focus on work.

Message routing

Telegram messages hit the correct agent based on which bot received them

Session management

Each conversation is a session with its own context window

Tool access

Web search, browser control, file read/write, shell commands, cron management

Memory

Long-term memory in markdown files, searchable via semantic search

Cron jobs

22+ scheduled tasks running on autopilot daily

Skills

255 skills.
21 categories.

Every skill is a markdown file following a strict protocol called FloatPrompt. Skills are not documentation — they're behavioral programs. When an agent loads a skill, it changes how the agent thinks, what tools it reaches for, and what it refuses to do.

Category

Skills

Shared agent skills

Browser automation, email triage, calendar, spawning, skill gap analysis

25

President (mdsbot)

CRM, revenue dashboards, coaching, newsletter, health tracking

38

Infrastructure (Q)

Gateway config, agent activation, health checks, updates, auth

10

Skill authoring (Napoleon)

Skill writing, persona design, credential management, workflow

7

Curriculum (Shifty)

Script authoring, lesson mapping, voice calibration, homework

11

Marketing (Mozi)

Hormozi frameworks, funnel architecture, voice guide

4

Content (Adley)

Hook engineering, shortform scripts, subject lines, tweet threads

13

Chief of Staff (Harmon)

Briefings, email triage, calendar, follow-ups, meeting prep

7

Total (21 categories)

255

FloatPrompt™

The methodology that governs how every skill, persona, and behavioral rule is written. Not a template — a design discipline with strict principles.

Behavioral programs, not documentation

Skills change how an agent acts, not just what it knows. Loading a skill rewires behavior.

STOP blocks set the frame

The first thing an agent reads in a skill constrains its behavior before it starts. Frame before execution.

Three archetypes

Reference (lookup), Behavioral (changes how you work), Decisional (changes how you decide). Every skill is one of these.

30-second scannable

If a skill can't be understood in 30 seconds, it's too long. Bulk goes to reference docs, not the main file.

Every field prevents drift

If you can't name the failure mode a config field prevents, delete it. No cargo-cult configuration.

Cold test everything

A new agent with no prior context should load the skill and operate correctly. No assumed familiarity.

The Recursive Loop

Agents improve their own tooling by using it. No human intervention. The more work they do, the better the skills get. It compounds.

01

Gap surfaces

An agent encounters a problem — wrong file path, missing command, capability gap — during normal work.

02

Report filed

The agent automatically files a gap report to Napoleon's inbox. No human intervention needed.

03

Skill updated

Napoleon reads the report and writes or updates the skill file.

04

Deployed

The updated skill gets wired into the agent's configuration.

05

Compound

Next time the agent loads, it has the fix. The more work agents do, the better the skills get.

Communication

Agents talk
to each other.

Four distinct channels depending on what the work requires. Async file handoffs, real-time session messages, single-shot subagents, and parallel coding teams. The right channel for the right task.

Inbox System

Every agent has an inbox folder. Any agent can drop a markdown file into another agent's inbox — a task, a gap report, a handoff brief. A file system watcher detects the new file and wakes the recipient agent via webhook. The agent processes the item, acts on it, and moves it to their log folder.

File system → webhook → agent

Direct Messaging

Agents can send messages directly into each other's active sessions. Used for real-time coordination when something can't wait for the inbox — a status update, a blocker, a question mid-task.

Session → session

Subagent Spawning

An agent can spawn another agent as a subagent — a one-shot task that runs, returns a result, and closes. Or a persistent session for ongoing collaboration. The parent agent yields its turn, waits for the result, then continues.

Parent yields → child runs → result returns

Claude Code Teams

For heavy coding work, agents can spawn parallel Claude Code CLI sessions — multiple headless workers each tackling a different part of the same codebase simultaneously. One researches, one builds, one writes tests, one reviews. Results merge back together.

Parallel workers → single merge

Memory

Four layers.
Nothing forgotten.

Institutional memory compounds over time. Every decision, every lesson, every mistake gets logged. Six months in, the agents know the operation better than most employees would.

01

Session Context

Short-term

Each conversation has a context window. When it fills up, the system compacts — summarizing older messages to free space while preserving key decisions and context. The agent keeps working without losing the thread.

02

Session Logs

Medium-term

Every significant session gets logged to a dated markdown file. These are searchable via semantic search using vector embeddings. When an agent needs to recall a past decision, it queries these logs.

03

Curated Memory

Long-term

A hand-maintained file of the most important decisions, lessons, patterns, and rules. Hard-won knowledge that should never be forgotten. Things like 'never do X because we tried it and here's what happened.'

04

Structured Database

Permanent

A hosted SQL database stores structured records — deep context documents, indexed session summaries, granular event logs, and active project tracking with status and metadata.

Database

A hosted SQL database accessible from both machines, storing everything that needs structure.

100

deep entries

Long-form context documents on major topics

515

session logs

Indexed conversation summaries

972

micro-logs

Granular event records: decisions, observations, file changes

17

active projects

Project tracking with status, owners, and metadata

Cron Jobs

22+ jobs.
Running without
supervision.

The system has a heartbeat. Scheduled jobs run whether anyone is at the keyboard or not. Health checks pass. Emails get triaged. Briefings get written. The human shows up to the good part.

Morning briefing

Rhythm

Daily, 7am

Calendar, email, and messages synthesized into a single daily brief. Delivered to the human before the first meeting.

Email triage

Rhythm

Daily, 8am

Inbox classified and routed. High-priority flagged. Routine handled. Noise filtered.

Follow-up tracking

Rhythm

Daily

Overdue commitments surfaced and escalated. Nothing falls through the cracks because a cron catches it first.

Health checks

Infrastructure

Multiple daily

Database integrity verified. Gateway status confirmed. Agent configurations validated. Silent failures surfaced before they compound.

Workspace maintenance

Infrastructure

Daily

File structure audits. Stale index detection. Naming convention enforcement. The workspace stays clean without anyone cleaning it.

Data collection

Research

Daily / weekly

Market intelligence gathered. Review platforms scanned. Content research filed. The agents know what's happening in the world.

This is not a chatbot. It's a workforce.

11 agents. 255 skills. 22+ cron jobs. Running daily across two machines and a phone.