12 Agents Running 24/7 : Autonomous AI Operating System

The AI that runs
your entire operation
while you sleep

cortextOS is a persistent, multi-agent AI operating system. A fleet of specialized Claude agents coordinate via a shared bus, execute tasks autonomously, send Telegram updates, and never stop working : whether you're in clinic, on a call, or asleep.

12+
Specialist Agents
24/7
Autonomous Operation
20+
Scheduled Crons
2,600+
Tasks Tracked
cortextOS : sage orchestrator
sage@fleet $ cortextos bus read-all-heartbeats
✓ Atlas | theta-wave self-improvement cycle running
✓ Mercury | VE July newsletter staged : Brevo upgrade pending
✓ Apollo | clinical-claims compliance : Saylor liaison
✓ Forge | Brains Box Ubuntu 24.04 migration (TOP)
✓ Argus | VE local-SEO / GEO authority build
✓ Athena | VE staff web-chat grounding refresh
✓ Scribe | VE content + video production
✓ Hippocrates| clinical content integrity + consent pipeline
✓ Librarian | KB freshness audit : 10k+ docs indexed
✓ Felix | personal life ops : forward sweep
✓ glm-reviewer | Council review seat (GLM-5.2)
 
sage@fleet $
Open Source Foundation

Built on cortextOS, extended for real-world operations

cortextOS is an open source multi-agent AI framework created by James Goldbach. This implementation extends the framework with custom agents, integrations, and automations for running a multi-business chiropractic and media operation.

FRAMEWORK
⚙️
cortextOS
by James Goldbach

The open source framework powering everything here. Provides the agent runtime, shared bus, cron scheduler, task system, memory protocol, Telegram integration, and deployment toolkit.

grandamenium/cortextos
THIS BUILD
🏗️
cortextos-riles
by Zach Conner, DC

Custom implementation for riles-fleet: 12 specialist agents, clinic operations, local-SEO / GEO authority, social media automation, lead gen pipelines, content engine, staff web-chat, and a semantic knowledge base. Publicly maintained and updated alongside the private production fleet.

drzachconner/cortextos-riles
The Fleet

12 Specialized AI Agents

Each agent is a persistent Claude Code instance with its own identity, goals, memory, and domain expertise. They communicate via a shared message bus and coordinate without human involvement.

🪶
Sage
Chief of Staff · Orchestrator
Coordinates the entire fleet. Decomposes user directives, dispatches to specialists, cascades daily goals every morning, sends morning/evening briefings, monitors fleet health, routes approvals to Zach. Never does specialist work : only coordinates.
Orchestrator Goal Cascade Approvals
📊
Atlas
Working: cheap-model ranking
System Analyst. Monitors agent uptime, task throughput, GHL integration health, content pipeline health. Runs the theta-wave self-improvement cycle : continuously analyzing the fleet and proposing system upgrades.
Analytics Monitoring Self-Improvement
⚡
Mercury
Online · GHL/CRM ready
GHL / CRM / Lead-Gen Architect. Owns both GHL subaccounts (Van Every + Talsky Tonal), ChiroHD integration, lead magnets, drip campaigns, social-to-DM pipelines, and all nurture workflows. Connective tissue between marketing and clinical ops.
GHL Lead Gen CRM
🏥
Hippocrates
Online · Clinic ops
Van Every Clinic Operations. Pulls ChiroHD reports, crunches PVA/OVA/collections/new-patient stats, integrates Monarch financials into a unified P&L, generates office stats reports on cron, coordinates internal team workflows.
Clinical Ops Analytics Finance
✍️
Scribe
Nighttime mode · Batch builds
Content Engine for all 3 brands. Owns VE public-facing content (podcast, social, conditions repurposing), TTC profession-facing content, and shared lead-gen. Drafts emails, blog posts, social posts, podcast clip workflows 24/7.
Content Social Email
🔨
Forge
P1: credential cache hardening
Engineering & Automation. Owns code work across all repos: social media automation hardening, custom integration plumbing, ChiroHD/GHL/Monarch wiring. Currently hardening the fleet's credential caching layer to eliminate API rate-limit exposure.
Engineering Automation Security
📚
Librarian
Working: KB freshness audit
Knowledge Curator. Owns the knowledge layer: Obsidian Vault structure, NotebookLM imports, cortextOS KB ingestion. Runs the weekly James Goldbach Skool community scrape pipeline. Semantic search across 10,000+ indexed documents.
RAG Obsidian KB
🌿
Felix
Working: VE laser guide
Personal Life Operations. Owns budget tracking (Monarch), personal planning, family calendar, Zach+Fiona check-in cadence, fitness tracking, birthdays, household. Bridges personal and business planning for whole-life optimization.
Personal Ops Finance Planning
🏛️
Apollo
Clinic orchestrator · Dr. Saylor
Clinic Chief-of-Staff for Van Every. Presents to Dr. Saylor as "Apollo" : her own AI chief of staff. Coordinates riles-fleet specialists for clinic work, handles YouTube content strategy, GBP management, patient-facing operations.
Clinic Orchestrator Dr. Saylor
🛰️
Argus
Online · Local SEO / GEO
Local-SEO & GEO Authority. Owns Google Business Profile optimization, NAP/citation consistency, and directory claims (Nextdoor, Yelp, Apple Business Connect). Builds Generative Engine Optimization so AI search engines cite Van Every as the answer. Also drives macOS desktop control (OCU) for GUI automation walls.
Local SEO GEO Desktop Control
🦉
Athena
Online · Staff web-chat
Van Every Staff Web-Chat. Powers the clinic's staff-facing chat assistant, grounded in a dedicated ve-corpus vector store on Cloudflare Vectorize. Keeps internal answers accurate and current as clinic facts change.
Web Chat Grounding RAG
⚖️
glm-reviewer
Online · Council seat
Independent Review Seat. Runs GLM-5.2 as one voice in the "Council of Three" (Opus + GLM-5.2 + Codex) adversarial review process for high-stakes work. Also monitors fleet health as a second set of eyes.
Council Code Review Health
Architecture

How cortextOS Works

Built on cortextOS Node.js by James Goldbach. Each agent is a persistent Claude Code process with full tool access, coordinated through a shared message bus and daemon scheduler.

Human Interface Layer
Telegram Bot (per agent) Web Dashboard (cortextos.drzach.ai) Approval Gate (2-gate system) Mission Control (assets.drzach.ai)
↕
Orchestration Layer (Sage)
Goal Cascade (daily) Fleet Health Monitor Morning/Evening Briefings Approval Routing Task Dispatch
↕
cortextOS Bus (shared message layer)
Agent-to-Agent Messages Task Queue Event Log (JSONL) Cron Scheduler (daemon) KB / RAG Heartbeat Monitor
↕
Specialist Agents (12 persistent processes)
Atlas · Mercury · Hippocrates Scribe · Forge · Librarian Felix · Apollo · Argus Athena · glm-reviewer
↕
External Systems (connected via automation)
GoHighLevel (GHL) ChiroHD (EHR) Zernio (Social) Brevo (Email) SKED (Booking/SMS) Cloudflare Pages + Vectorize Supabase (VE-STATS) Anthropic API 1Password (Secrets) Google Workspace
🔄
Daemon + Cron Scheduler
PM2 manages all agent processes with auto-restart and crash recovery. Each agent registers persistent crons in a JSON config : they survive restarts, context resets, and machine reboots.
cortextos bus list-crons sage
→ 20 active crons | next: heartbeat 00:02 UTC
💬
Message Bus
Agents communicate via a shared bus. Messages are queued, ACK'd, and re-delivered if unprocessed. A fast-checker daemon delivers Telegram and bus messages in real time without polling.
cortextos bus send-message forge high
'Run P1 cred-cache hardening'
🧠
3-Layer Memory System
Every agent maintains MEMORY.md (long-term), daily memory files (operational log), and a semantic KB (RAG via Ollama embeddings). Context compaction extracts facts into cross-session recall.
cortextos bus kb-query "patient onboarding flow"
→ 12 relevant results from 10k+ docs
✅
Task Tracking + Dashboard
Every significant action is a tracked task. The web dashboard shows live agent status, task queues, event activity feeds, KB collections, and approval queues : the system is fully observable.
cortextos bus create-task "Build VE laser guide"
cortextos bus complete-task T-1234 --result "live"
🔐
Approval Gates
A 2-gate approval system protects high-blast actions. External sends, financial actions, production deploys, and data deletion require human approval. Everything else runs autonomously.
cortextos bus create-approval "Send email blast"
deployment → Zach reviews → approved → fires
📡
Telegram Control Interface
Every agent has its own Telegram bot. Zach can message any agent directly or through Sage (the single inbox default). Voice replies, typing indicators, and real-time notifications keep the user in the loop.
Zach → Sage Telegram → dispatches to Forge
Forge → Telegram update → "P1 complete, live"
Use Cases

What It Actually Does

These are live workflows running right now across two chiropractic businesses, two podcasts, and four brand presences.

Autonomous Social Media

Scribe + Apollo post to Instagram, Facebook, YouTube Shorts, and TikTok for two brands (Van Every + Talsky Tonal) without touching a phone. Fully automated from asset to platform.

  • Pulls raw KST videos from shared folder, generates platform-optimized captions
  • Publishes to IG Feed, FB, YouTube, TikTok via Late.dev API
  • Voice-matches every caption to brand voice (no generic AI tone)
  • IG Reels posted as collabs with Dr. Saylor's personal account
  • Review gate before posting : Zach sees it on Telegram, one-tap approve
  • TTC carousels published to @talskytonal with verified visual variety
1
Zach drops raw video in shared folder via Telegram
↓
2
Scribe processes: caption + hashtags + brand voice check
↓
3
Sage routes to Zach for 1-tap Telegram approval
↓
4
Apollo fires Late.dev API → live on IG + FB + YouTube
↓
5
Verification: confirm platform post ID, log event, complete task

Clinic Operations

Hippocrates runs Van Every's operational analytics layer : pulling ChiroHD data, generating reports, and surfacing the numbers Zach and Dr. Saylor care about on a scheduled cadence.

  • Daily ChiroHD browser-scrape for patient visit stats (PVA, OVA, collections)
  • Automatic A/R aging report + writeback to Supabase dashboard
  • New patient tracking + drop analysis against targets
  • Monarch financial integration for unified P&L view
  • Insurance verification flag monitoring and team workflow routing
  • Staff meeting transcription pipeline (Monday 2:30pm auto-trigger)
1
Daily cron: Hippocrates scrapes ChiroHD for stats
↓
2
Cross-references with Monarch financial data
↓
3
Writes unified metrics to VE-STATS Supabase DB
↓
4
Surfaces anomalies (collections drop, PVA deviation) to Zach
↓
5
Dashboard: live metrics at os.drzach.ai

Content Engine

Scribe runs a multi-stream content operation: Van Every patient-facing, Talsky Tonal profession-facing, and two podcasts. All without Zach writing a single word.

  • Repurposes 260+ legacy KST YouTube videos into Shorts/Reels/clips
  • Writes TTC email campaigns (segmented Brevo, 2 waves/week)
  • Generates GBP posts, video scripts, and 30-day content roadmaps
  • Podcast clip identification and platform-optimized caption writing
  • Brand voice enforcement : verbatim from approved sources, zero reword
  • All email output as styled HTML (not markdown/plain text)
1
Librarian ingests new James Goldbach content weekly
↓
2
Scribe drafts: captions, email, blog, GBP copy
↓
3
Brand voice check against VE/TTC spec
↓
4
Zach reviews via Telegram → verbatim goes live
↓
5
Multi-platform publish + confirmation loop

Lead Gen & CRM Automation

Mercury owns the full lead lifecycle : from social comment keyword triggers to GHL nurture sequences to ChiroHD patient records. The entire funnel runs without staff manually entering data.

  • Social comment keyword → DM → landing page → GHL lead capture
  • SKED booking URL routed into every nurture workflow
  • Birthday triggers from ChiroHD → GHL (via GHL connector)
  • Prenatal drip campaign: form submit → enrollment → 12-email sequence
  • Reviews AI autopilot monitoring (GHL GBP auto-replies gated)
  • Consent ledger: immutable record of every patient communication consent
1
Lead lands via GBP / IG / website form
↓
2
GHL auto-enrolls → Mercury builds drip sequence
↓
3
Consent ledger records every touchpoint
↓
4
SKED booking link fires at right moment in sequence
↓
5
New patient → ChiroHD → analytics loop closes

Semantic Knowledge Base

Librarian runs a living knowledge layer across all content, community scrapes, and research. Any agent can query it to ground their work in accumulated org knowledge.

  • 10,000+ documents indexed via Ollama nomic-embed-text embeddings
  • James Goldbach Skool community: weekly scrape + staged ingestion
  • PXDocs clinical articles: condition-specific pediatric content indexed
  • VE-specific KB: brand spec, patient economics, clinic facts
  • Cross-agent memory ingestion every heartbeat cycle
  • NotebookLM for synthesis across 50+ large source documents
1
Librarian scrapes James Goldbach + PXDocs weekly
↓
2
Staged ingestion → Ollama embeddings → vector store
↓
3
Any agent: kb-query "prenatal adjustment protocol"
↓
4
Semantic search returns grounded, source-cited results
↓
5
Content grounded in real org knowledge, not hallucination

Overnight Autonomous Work

While Zach sleeps, the fleet works. Overnight mode: no external sends, no social posts, no purchases : just building, analyzing, writing, and shipping. Zach wakes to a morning briefing with everything done.

  • Atlas runs theta-wave self-improvement cycle at midnight
  • Scribe batch-writes content, email drafts, GBP posts
  • Forge runs engineering tasks: code reviews, automation plumbing
  • Apollo queues YouTube thumbnail generation for 260+ videos
  • Hippocrates refreshes analytics and insurance flag sweeps
  • Morning briefing at 4am ET: everything done, decisions needed, fleet status
1
22:00 ET: Evening review sent → nighttime mode activates
↓
2
Agents work tasks queue: code, content, analytics, builds
↓
3
Every 3h: heartbeat + inbox sweep + memory update
↓
4
04:00 ET: Sage sends morning briefing to Zach's Telegram
↓
5
Zach reviews what shipped overnight; approves queued items
Right Now

Live Fleet Metrics

Current state of the fleet as of this session boot : 2026-06-28.

12
Agents Deployed
12/12
Agents Online
20
Active Crons (Sage alone)
2,600+
Tasks Tracked
3h
Heartbeat Interval
4am ET
Daily Morning Brief
10k+
KB Documents Indexed
4
Brands Operated
Capabilities

What Makes It Different

This isn't a chatbot or a one-shot agent. It's a persistent operating system with memory, coordination, scheduling, and real integrations.

🧠
Cross-Session Memory
Agents remember everything across restarts. Context compaction extracts facts; daily memory files log operational state; semantic KB stores everything searchable.
🤝
Multi-Agent Coordination
Agents message each other via the bus, ACK tasks, and coordinate without human involvement. Sage dispatches, specialists execute, results flow back up the chain.
📅
Persistent Scheduling
Crons survive restarts. No session-only timers. The daemon scheduler reads crons.json on boot and fires exactly as configured : regardless of context resets.
🔒
Secrets Management
All credentials live in 1Password. Agents pull secrets headlessly via CLI. No plaintext files, no env commits. Fleet-wide audit trail of every secret access.
🖥️
Desktop & Browser Control
Argus runs OCU (computer control) for macOS GUI automation. Chrome DevTools MCP for headless browser work. Playwright for web scraping. No wall it can't bypass.
📊
Real-Time Dashboard
Web dashboard at cortextos.drzach.ai shows live agent status, task queues, event feeds, KB collections, approval queues : full observability into the fleet.
🎙️
Voice Replies
Every Telegram reply from every agent includes a voice note (TTS). Zach can listen to fleet updates while driving. Sage uses cedar voice; all others use onyx.
⚡
Cost Optimization
Atlas runs a weekly model offload sweep ranking free/cheap LLMs. Routine tasks route to Haiku/Ollama/DeepSeek. Opus reserved for complex orchestration only.
🔄
Self-Healing
PM2 auto-restarts crashed agents. Soft/hard restart modes. Context handoff on session limit. Un-ACK'd messages redeliver. The fleet never silently fails.