Phalanx logoPhalanx

AGENT FLEET08 SPECIALIZED ROLESACCOUNTABILITY ENFORCED

Distinct roles.Distinct prompts.Cognitive diversity by design.

Each agent has a defined role, a custom system prompt, and a clear accountability boundary.

None of them can do another's job — the architecture forces collaboration. This is the opposite of a single LLM call wearing seven hats.

Read the roster ↓See the public ledger

THE ROSTER / I–VII

Eight specialists. Seven accountability boundaries.

Each row is one agent. The CAN / CANNOT columns name the architectural boundary — what the role is allowed to do, and what it's explicitly not. Ordered top-to-bottom in the flow of a decision cycle.

  1. I.

    CHIEF

    Orchestrator · routes the fleet

    Routes tasks across the agent fleet. Sets the morning agenda, escalates issues, summarizes the system to the operator at end of day. Has authority to gate trades.

    CAN

    • Route work across the other six agents
    • Set the morning agenda + EOD summary
    • Gate trades before they reach execution

    CANNOT

    • Place orders directly — the engines execute
    • Override the evaluator's audit verdicts
    HANDS OFF →Operator (EOD) · Trader (gated approvals)
  2. II.

    SCANNER

    Eyes on the tape · candidate generation

    Continuously scans the market for flow patterns, unusual options activity, premarket movers. Surfaces candidates; does not decide.

    CAN

    • Watch real-time order flow + options activity
    • Surface candidate signals with a confidence tag

    CANNOT

    • Issue a verdict — surfacing is not deciding
    • Trigger execution
    HANDS OFF →Researcher (for deep dive) · Chief (for routing)
  3. III.

    ANALYST

    Regime + macro · daily recommendation

    Reads the macro picture daily — volatility regime, sector rotation, dollar strength, credit spreads. Produces the morning regime recommendation that gates every engine.

    CAN

    • Read macro data + produce regime recommendation
    • Gate every engine via the regime file

    CANNOT

    • Place trades
    • Override the chief's routing decisions
    HANDS OFF →Chief (regime file) · Trader (size + stop adjustments)
  4. IV.

    RESEARCHER

    Deep dives on demand · dossier output

    When a candidate emerges, the researcher pulls earnings, filings, peer comps, sentiment, and prior performance. Outputs a dossier, not a verdict.

    CAN

    • Pull filings, sentiment, peer comps on demand
    • Output a structured dossier per candidate

    CANNOT

    • Issue a verdict — dossiers inform, they do not decide
    • Trigger execution
    HANDS OFF →Trader (for sizing decisions) · Chief (for context)
  5. V.

    TRADER

    Entry · sizing · stops

    Decides on entry, position size, and exit conditions for high-conviction trades. Operates only inside the regime constraints set by the analyst and approved by the chief.

    CAN

    • Decide entry price, position size, exit conditions
    • Submit order specs to the execution engine

    CANNOT

    • Operate outside the regime gate
    • Bypass the chief's gate or the consensus vote
    HANDS OFF →Execution engine (order placement) · Ledger (signal log)
  6. VI.

    ENGINEER

    Infrastructure watchdog · health + paging

    Monitors the trading engines, gateways, data feeds, and cron schedules. Pages the operator when anything degrades. The reason the rest of the fleet keeps running.

    CAN

    • Monitor engines + gateways + data feeds + crons
    • Page the operator when SLOs degrade

    CANNOT

    • Place or modify trades
    • Edit strategy parameters
    HANDS OFF →Operator (alerts) · Chief (system health summary)
  7. VII.

    EVALUATOR

    Independent auditor · isolated infrastructure

    Runs on isolated infrastructure. Grades every other agent at end of day. Reports only to the operator. Has no execution authority. The reason this is governance, not just monitoring.

    CAN

    • Audit every signal pre-trade
    • Grade every other agent nightly
    • Report verdicts to the operator

    CANNOT

    • Place, cancel, or modify trades
    • Be read, forked, or overridden by any other agent
    HANDS OFF →Operator only · no agent can read these verdicts

    → INDEPENDENT · ISOLATED · REPORTS ONLY TO OPERATOR

WHY THIS MATTERS

An auditor whose job depends on agreeing isn't an auditor.

In most multi-agent systems, every agent runs in the same process, with the same model, on the same prompt template — calling itself with different system messages. That's not governance. The evaluator's incentives are entangled with the system it's supposed to grade.

THE SEPARATION

In Phalanx, the evaluator runs on a separate gateway, with a separate model, on isolated infrastructure. It cannot place trades. It grades the others nightly. It reports only to the operator.

This separation is the architectural equivalent of an external auditor — and it's the single most load-bearing piece of the system's credibility.

WHAT GOVERNANCE LOOKS LIKE HERE
  • GATEWAYSeparate · isolated from the agent fleet
  • MODELDifferent foundation model from any in the consensus gate
  • AUTHORITYRead-only · cannot place, cancel, or modify trades
  • REPORTSOperator only · no agent can read these verdicts
→ SEEEvery evaluator verdict on the public ledger

LAYER II OF V

Every decision these agents make is on the record.

See the public ledger →
The agents — Phalanx