How a 32-agent hedge fund decides — and when it refuses to.

The bedrock philosophy

Every agent reads context/philosophy.md at boot. The doctrine is short and load-bearing:

1. Survival before return. A blown account cannot compound. Capital preservation is the prerequisite for every other goal.
2. Uncertainty equals WAIT. The default action is not action. Confidence below threshold means hold, not size down.
3. The market is the boss. Stance follows reality. AGGRESSIVE/NORMAL/DEFENSIVE/LOCKDOWN is not a knob the operator sets; it is the system's read of conditions.
4. Mistakes are journaled, not punished. Every miss writes to context/journal/. The SelfOptimizer rewrites tomorrow from what failed today.
5. Emergency is supreme. The watchdog L4 HALT can override every other agent, including the CEO. Failure is fail-closed.

The org chart

Five departments + a research desk + a CEO. Each agent owns one decision; no agent owns the whole trade.

• CEO — Sets stance, holds the weekend, declares code-red, responds to broker margin events.
• Macro Officer — Reads geopolitics, central banks, broker mail. Outputs an independent stance: RISK_ON / RISK_OFF / MIXED / CRISIS.
• Risk Department — Emergency, PanicControl, RiskGate, ExposureManager, AdaptiveParams. The veto layer.
• Strategy Department — Strategist, StrategyBank, StrategySwitcher, Confluence, Backtest, MeanReversion, Patterns, and tier-1–4 expansion agents.
• Execution Department — TradesManager, Coordinator, FactChecker, DoubleChecker.
• Intelligence Department — News, Calendar, Journal, SelfOptimizer, and three Report Writers (hourly, daily, weekly).

The 10-step lifecycle

Every agent runs the same 10-step loop on every wake-up: read philosophy → read its own context → read the active stance → read the message bus → form a hypothesis → consult its skills → emit a proposal → sign for the FactChecker → wait for DoubleChecker → journal the result. No agent skips a step.

The three gates

Every decision passes three independent gates before reaching the EA:

1. Risk Gate — Has stance + size + drawdown headroom + correlation exposure within bounds?
2. FactChecker — Are the inputs that justify this decision still true at the moment of signing?
3. DoubleChecker — Does a second, blind agent reach the same conclusion?

A trade that any gate rejects becomes a journal entry, not a position.

The emergency reflex

An independent watchdog runs outside the agent graph. It monitors drawdown, margin utilization, broker connectivity, and stop-loss invariants. If any invariant breaks, it issues an L4 HALT that closes positions, locks the EA, and pages the operator. The CEO cannot countermand a watchdog halt.

Compounding intelligence

Every night, a Self-Optimizer agent reads the day's context/journal/, scores what worked, retires what didn't, and queues experiments for tomorrow. The system gets smarter without changing operators. The audit trail is markdown — you can read why a setup was retired six months from now.