Autopilot Mode

Autopilot mode enables agents to operate autonomously as a self-running company. Agents propose work, the system enforces resource constraints (cap gates), approved proposals become executable missions, and events trigger new proposals — creating a closed-loop cycle with no human intervention required.

Architecture

Proposal → Cap Gate → Approval → Mission → Execution → Event → Trigger/Reaction → Proposal

Three components work together:

• AgentZero (VPS) — The brain. Runs the agent loops, executes missions, enforces cap gates.
• Supabase — State layer. Stores proposals, missions, events. Provides real-time subscriptions.
• Next.js/Vercel (separate repo) — Dashboard. Read-only view + lightweight control (approve/reject).

Quick Start

1. Enable Autopilot

Add to agentzero.toml:

[autopilot]
enabled = true
supabase_url = "https://xxx.supabase.co"
supabase_service_role_key = "eyJ..."
max_daily_spend_cents = 500
max_concurrent_missions = 5
stale_threshold_minutes = 30
reaction_matrix_path = "reactions.json"

2. Define Agents

[[agents]]
name = "editor"
system_prompt = "You are the Editor-in-Chief. Propose blog topics and review content."

[[agents]]
name = "writer"
system_prompt = "You are a Content Writer. Write posts from research briefs."

3. Add Triggers

[[autopilot.triggers]]
name = "periodic_topic_proposal"
condition = { type = "cron", schedule = "0 */6 * * *" }
action = { type = "propose_task", agent = "editor", prompt = "Propose a new blog topic." }
cooldown_secs = 21600

4. Configure Reactions

Create reactions.json:

[
  {
    "source_agent": "editor",
    "event_pattern": "proposal.approved",
    "target_agent": "writer",
    "action": "begin_writing",
    "probability": 1.0,
    "cooldown_secs": 60
  }
]

5. Apply Supabase Schema

Run supabase/migrations/001_autopilot_schema.sql against your Supabase project.

Core Concepts

Proposals

Agents create proposals for work to be done. Each proposal has a type, priority, and estimated cost:

Field	Description
proposal_type	content_idea, task_request, resource_request, system_change
priority	low, medium, high, critical
estimated_cost_microdollars	Cost estimate (1 cent = 10,000 microdollars)
status	pending → approved / rejected → executed

Cap Gates

Cap gates enforce resource constraints before a proposal is approved:

Constraint	Default	Description
max_daily_spend_cents	500	Maximum daily spend across all agents
max_concurrent_missions	5	Maximum missions running simultaneously
max_proposals_per_hour	20	Proposal rate limit
max_missions_per_agent_per_day	10	Per-agent mission cap

A proposal that violates any constraint is automatically rejected with a reason.

Missions

When a proposal is approved, it becomes a mission with executable steps:

Field	Description
status	pending → in_progress → completed / failed / stalled
steps	Ordered list of MissionStep with individual status tracking
heartbeat_at	Last heartbeat timestamp (used for stale detection)
deadline	Optional deadline

Triggers

Triggers create new proposals when conditions are met:

Condition Type	Description
event_match	Fire when a specific event type is emitted
cron	Fire on a cron schedule (e.g., 0 */6 * * *)
metric_threshold	Fire when a metric exceeds a threshold

All triggers support cooldown_secs to prevent rapid re-firing.

Reaction Matrix

The reaction matrix defines probabilistic inter-agent interactions. When agent A emits event X, agent B has probability P of proposing action Y.

{
  "source_agent": "writer",
  "event_pattern": "mission.completed",
  "target_agent": "social_media",
  "action": "propose_social_post",
  "probability": 0.9,
  "cooldown_secs": 600
}

Event patterns support wildcards: * matches any event, content.* matches all content events.

Stale Recovery

Missions with heartbeat_at older than stale_threshold_minutes are automatically marked as stalled, and a mission.stalled event is emitted. This prevents stuck missions from consuming cap gate capacity.

Company Templates

Three pre-built templates are available in templates/:

Content Agency

6 agents: editor, researcher, writer, seo, social_media, analyst. Agents autonomously propose topics, research, write, optimize, publish, and promote content.

Dev Agency

6 agents: pm, architect, coder, reviewer, devops, support. Agents triage issues, design solutions, write code, review PRs, deploy, and handle support.

SaaS Product

6 agents: product, engineer, qa, support, growth, ops. Agents manage product development, testing, customer support, growth marketing, and operations.

Each template includes an agentzero.toml with agent definitions and triggers, plus a reactions.json with the reaction matrix.

Autopilot Tools

Four tools are available to agents when autopilot is enabled:

Tool	Description
proposal_create	Create a new proposal with type, priority, and cost estimate
proposal_vote	Approve or reject a proposal (auto-creates mission on approval)
mission_status	Query one or all missions, optionally filtered by status
trigger_fire	Manually fire a trigger by ID (for testing or agent-initiated reactions)

Human-in-the-Loop

Autopilot and human interaction work together:

• Dashboard approval — Proposals can be auto-approved by cap gates OR manually approved/rejected from the dashboard via /v1/autopilot/proposals/:id/approve.
• Chat — Agents still respond to chat via channels (Telegram, Discord, Slack). Human messages can influence agent behavior alongside autonomous operations.
• Trigger control — Triggers can be enabled/disabled from the dashboard without restarting.
• Emergency stop — The existing estop mechanism halts all autonomous operations immediately.

Gateway Endpoints

When autopilot is enabled, the gateway exposes additional REST endpoints for dashboard control:

Method	Path	Description
GET	/v1/autopilot/proposals	List proposals (paginated)
POST	/v1/autopilot/proposals/:id/approve	Approve a proposal
POST	/v1/autopilot/proposals/:id/reject	Reject a proposal
GET	/v1/autopilot/missions	List missions
GET	/v1/autopilot/missions/:id	Mission detail with steps
GET	/v1/autopilot/triggers	List triggers
POST	/v1/autopilot/triggers/:id/toggle	Enable/disable a trigger
GET	/v1/autopilot/stats	Daily spend, mission counts, agent activity

Supabase Schema

The SQL migration (supabase/migrations/001_autopilot_schema.sql) creates 8 tables:

Table	Description
proposals	Proposal records with status tracking
missions	Mission records with heartbeat and deadline
mission_steps	Individual mission steps with status
events	Event log for triggers and audit
triggers	Trigger rule configuration
content	Published content (blog posts, etc.)
agent_activity	Agent activity log
cap_gate_ledger	Cost tracking for cap gate enforcement

RLS policies grant full access to the service role (AgentZero VPS), read-only to anon (dashboard), and public read on published content only.