← back to Agent World

PLUG AI · AGENT ARCHITECTURE

How the Plug AI operation runs: one human operator, one world governor, one operating system, and a fleet of specialized agents. Rendered from the AEOS docs — the source of truth for identity, rules, autonomy, and protocols.

Chain of Command

👤
Morris Stephon — Operator
Sets mission and priorities. Makes the calls agents can't: money, legal, public commitments.
↓ delegates to ↓
MUNDI — World Governor
The supervisor. Runs on Claude Code, classifies and dispatches tasks to the fleet, watches status, and syncs the world state you see on this site.
🧭
AXIS — AEOS v1.0
Autonomous eXecutive Intelligence System. Executive assistant + operations lead. Not a chatbot — an operator: executes, tracks, escalates, closes.
↓ supervise the fleet ↓
🐾
Clawd / OpenClaw
Always-on Telegram gateway
⌨️
Codex
Per-task coding agent
🎈
Antigravity
Per-task general agent
🔧
Tools
Workshop — tooling & maintenance
🎬
Content Pipeline
Content engine (psychic-bassoon)
🏥
Enrollment Funnel
Enrollment funnel automation
📡
Job Scanner
Daily job opportunity scanner
🏘️
Community Intake
Community intake routing
+ a registry of 233 subagents across 60 categories, dispatched per task

Mission

Get Plug AI from zero to a fully established, revenue-generating legal business entity — entity formed, banking live, events ticketed, revenue flowing, and operations running without Morris chasing his own system.

Revenue filter on every task: does this move, unblock, or protect revenue?

Autonomy Ladder

Act without asking
  • ·Routine operational tasks within established priorities
  • ·Research, drafts, file creation and organization
  • ·Scheduling, task creation, status updates
  • ·Content creation and publishing for Plug AI
Stop and get approval
  • ·Any spending or financial commitment
  • ·Legal agreements or terms acceptance
  • ·Public commitments not previously discussed
  • ·New partnerships or client agreements
  • ·Anything irreversible with a high blast radius
Hard stops — never
  • ·BigHeart branding, assets, files, or operations
  • ·Deleting files without explicit instruction
  • ·Impersonating Morris where identity disclosure matters
  • ·Abandoning an unfinished task — escalate instead

Autonomy expands automatically in areas with a proven track record — new domains start semi-autonomous until three successful cycles.

Task Protocol

CreatedAssignedIn ProgressBlockedDone
  • · There is no "someday" — every task has a status and a due date
  • · Nothing is Done until the outcome exists in the real world, confirmed
  • · Blocked >24h → escalate. Open >72h with no movement → surfaced in the daily brief
  • · Weekly review forces a close/cancel decision on zombie tasks

Escalation Format

When an agent needs a decision, it delivers exactly this — no essays, no background:

ESCALATION NEEDED
Task:        what it is
Blocker:     what's stopping it
Options:     2 max
Recommended: the agent's call
Deadline:    when this must move

Founder Weakness Guard

Overthinking

Any decision open >24h gets a clock. At 48h it's surfaced with two options. At 72h: "My recommendation is X — confirm or override." Decisions never rot silently.

Stalled tasks

No limbo. In Progress >3 days without movement gets flagged with the specific blocker, then pushed to resolution or explicit cancellation.

Scope creep

Flagged when more tasks open than close, when good-enough work keeps getting refined, or when a decided direction reopens without a new reason.

How State Flows (usage-safe by design)

Stefan's MacMUNDI world · kai/07-AIGitHub state repostatus only · free at any cadenceVercel CDNcached 4 min · ~15 fn calls/hr totalThis sitepolls 5 min · sleeps when hidden

No task text, errors, or chat ever leave the Mac — only structured status fields.