Solo-built prop trading firm with 3 production agents orchestrating live multi-broker execution. Running on FundedNext, E8, and LiteFinance since 2026-05.

Prop trading at scale requires more than a strategy. You need an ops layer that survives broker quirks (hedging-account OCO behavior differs from netting), withstands wine-MT5 fragility (terminals die without DISPLAY+XAUTHORITY), coordinates multiple agents without one stepping on another, and enforces kill rules that the model itself can't override. Off-the-shelf agent frameworks don't cover any of this. I built the stack from scratch.

                     ┌─────────────────┐
                     │   DeskHead      │  (cogito-2.1:671b)
                     │   playbook +    │
                     │   daily diary   │
                     └────────┬────────┘
                              │
        ┌─────────────────────┼─────────────────────┐
        ▼                     ▼                     ▼
  ┌──────────┐         ┌──────────┐         ┌──────────────┐
  │    QD    │         │Researcher│         │  HaltActor   │
  │  exec    │         │ defend + │         │  governance  │
  │ engineer │         │  extend  │         │ -1.5σ kill   │
  └────┬─────┘         └────┬─────┘         └──────┬───────┘
       │                    │                      │
       ▼                    ▼                      ▼
  ┌──────────────────────────────────────────────────┐
  │           paperclip-stonedesk  (fork)            │
  │  planexec adapter · phantom-XML strip · OpenCode │
  └─────────────────────────┬────────────────────────┘
                            │
                            ▼
  ┌──────────────────────────────────────────────────┐
  │             nautilus-mt5-journal                 │
  │  TRB strategies · OCO via TRADE_ACTION_SLTP      │
  └─────────────────────────┬────────────────────────┘
                            │
        ┌───────────────────┼───────────────────┐
        ▼                   ▼                   ▼
  ┌──────────┐        ┌──────────┐        ┌──────────┐
  │ FundedNext│       │   E8     │        │LiteFinance│
  │  MT5+EA   │       │ MT5+EA   │        │  MT5+EA   │
  └──────────┘        └──────────┘        └──────────┘

• 3-agent stack: DeskHead (playbook owner, daily diary), QD (live exec engineer, MT5 plumbing, 95% execution-match KPI), Researcher (defend + extend + ρ/cov + time-aware + 11-item self-validation checklist).
• paperclip-stonedesk fork with planexec adapter, phantom-XML strip defense, OpenCode local adapter patched to Ollama-only.
• HaltActor governance: state file at /var/lib/nautilus-strategy/<book>/halt_state.json, -1.5σ kill rule, operator pause is BUSINESS only — trading halts owned by HaltActor + DeskHead.
• Multi-broker execution via file-bridge EA architecture — see MT5 bridge case study.

Planner-executor adapter — format-json injection

// adapter forces structured output, kills XML drift
const planResponse = await ollama.chat({
  model: 'cogito-2.1:671b-cloud',
  format: 'json',
  messages: [
    { role: 'system', content: PLANNER_FEW_SHOT },
    { role: 'user',   content: task.prompt },
  ],
});
const plan = JSON.parse(planResponse.message.content);
// 30% XML rate → 0% under format:json + few-shot (G1 verdict 2026-06-09)

phantom-XML strip — defense layer

// strip-server/src/lib/phantom-strip.ts
const PHANTOM_TAGS = ['thinking', 'execution', 'reasoning', 'inner_monologue'];
export function stripPhantom(raw: string): { cleaned: string; stripped: Tag[] } {
  const stripped: Tag[] = [];
  let cleaned = raw;
  for (const tag of PHANTOM_TAGS) {
    const re = new RegExp(`<${tag}>[\\s\\S]*?</${tag}>`, 'gi');
    if (re.test(cleaned)) {
      stripped.push({ tag, reason: 'phantom-XML wrapper not in plan schema' });
      cleaned = cleaned.replace(re, '');
    }
  }
  return { cleaned, stripped };
}

HaltActor — state file pattern

# strategies/trb_london_range.py
HALT_STATE_PATH = f"/var/lib/nautilus-strategy/{book}/halt_state.json"
def is_halted():
    try:
        with open(HALT_STATE_PATH) as f:
            return json.load(f).get("halted", False)
    except FileNotFoundError:
        return False
# strategy code never decides halts; HaltActor + DeskHead own it

• Phantom-XML defense reduced cogito-2.1 zero-tool sessions from 57.5% to 0% on G1 smoke.
• OCO TRADE_ACTION_SLTP fix eliminated naked positions on hedging brokers (STO-49 critical fixed nautilus-mt5-journal@dc46f36).
• Auto-PATCH from EVIDENCE: contract closed the "claim-without-PATCH" loop where agents narrated done without firing API calls.