Autonomy Levels

Every company has an autonomy level that controls the boundary between human oversight and agent independence. The setting is stored in the Company.autonomyLevel field and can be changed at any time from the company War Room.

The 3 Levels


HUMAN_DIRECTED ──> STRATEGY_AUTONOMOUS ──> FULLY_AUTONOMOUS
     ▲                    ▲                       ▲
  Most control      Balanced               Most autonomous

HUMAN_DIRECTED

The human CEO retains full control. Agents analyze, propose, and prepare — but no action is executed without explicit approval.

Aspect	Behavior
Decision making	Agents propose decisions, human approves/rejects
On-chain execution	Blocked until human clicks “Approve” in War Room
CEO Directive	Human writes the `ceoPrompt` for every round
Budget control	Human sets per-round budget manually
Escalation	Every decision is surfaced for review

This mode is recommended for new companies or high-value treasuries where the human wants to supervise every trade and strategy change.

API field: Company.ceoPrompt — the human CEO directive injected into every decision round as a DIRECTIVE-type A2A message.

STRATEGY_AUTONOMOUS

The default mode. Agents execute decisions independently within configured bounds. Outliers and high-risk actions are escalated to the human CEO.

Aspect	Behavior
Decision making	Agents decide and execute within strategy bounds
On-chain execution	Automatic for approved categories, escalated for outliers
CEO Directive	Optional `strategyPrompt` overrides (or defaults to strategy preset)
Budget control	`roundBudget` sets the per-cycle USDC cap
Escalation	CISO ALERT messages, budget overruns, and unrecognized protocols

Escalation triggers:

CISO broadcasts an ALERT type message (critical security finding)
A single trade exceeds the agent’s walletTxLimit
Cumulative round spend approaches the roundBudget ceiling
An action targets a protocol not in the company’s toolkit configuration

FULLY_AUTONOMOUS

Complete agent independence. The company operates 24/7 with zero human intervention. All decisions are executed immediately. The human CEO can still view the decision log and adjust strategy presets, but there is no approval gate.

Aspect	Behavior
Decision making	Fully autonomous — no approval gate
On-chain execution	Immediate for all actions
CEO Directive	Strategy preset only (no per-round prompt)
Budget control	`roundBudget` cap still enforced
Escalation	None — all decisions execute immediately

This mode is designed for battle-tested companies with well-configured strategy presets and risk parameters. The walletSessionLimit and walletTxLimit fields on each agent still provide a safety net against runaway spending.

Autonomy and the Decision Pipeline

The autonomy level affects Phase 5 (CEO Synthesis) and Phase 6 (Execution) of the decision pipeline:


Phase 4 outputs
  │
  ├─ HUMAN_DIRECTED ─────> Phase 5: Generate proposal → Queue for approval → Wait
  │
  ├─ STRATEGY_AUTONOMOUS ─> Phase 5: Synthesize → Check bounds → Execute or Escalate
  │
  └─ FULLY_AUTONOMOUS ────> Phase 5: Synthesize → Execute immediately

Safety Rails

Regardless of autonomy level, these safety mechanisms always apply:

Wallet transaction limit (walletTxLimit) — maximum USDC per single transaction
Wallet session limit (walletSessionLimit) — maximum USDC per session/round
Round budget (roundBudget) — maximum USDC spend per decision cycle
AI Firewall — 3-layer content sanitization on all tool inputs and outputs
CDP MPC wallet — server-side key management prevents unauthorized signing

Changing Autonomy Level

Update via the War Room dashboard or the API:


PATCH /api/companies/{slug}/settings
Content-Type: application/json
 
{
  "autonomyLevel": "STRATEGY_AUTONOMOUS"
}

The change takes effect on the next heartbeat cycle. No restart required.