Source Watchers
Public news, calendars, and narrative signals. No broker keys. No execution rights.
AUTONOMOUS AGENT TRADING SYSTEM // CONCEPT BRIEF
Infrastructure with money attached.
A research architecture for market-aware agents: Python watches continuously, Hermes thinks selectively, risk policy blocks reckless behavior, and every meaningful action leaves an audit trail.
READ THIS FIRST
This is not a profit claim. It is a safety-shaped system design.
Futures are leveraged and dangerous. The goal here is not magic alpha or a button that prints money. The goal is a redeployable research skeleton that defaults to paper trading, isolates authority, records decisions, and makes dangerous states hard to reach by accident.
SERVICE MAP
Small containers. Narrow permissions. Clear handoffs.
Market Watchers
Price movement, volatility, volume, VWAP, trend shifts, and session context.
Event Router
Filters noise and decides when agent reasoning is worth the cost.
Hermes Researcher
Creates briefings from selected context. Observes; does not trade.
Hermes Trader
Drafts structured trade proposals. Proposes; does not execute.
Risk Governor
Validates proposals against hard policy. Blocks uncertainty by default.
Execution Bot
Paper broker first. Live mode requires explicit, current authorization and hard acknowledgments.
Audit Log
Records proposals, blocks, approvals, executions, errors, and kill-switch events.
CONTEXT MODEL
The agent thinks from bounded memory, not vibes.
Nyx's own operating context is a useful model: compact standing instructions, selective long-term memory, session recall, private documents, and fresh tool output. The trading stack uses the same principle: feed agents the right context at the right time, not everything all at once.
context_stack = {
standing_orders: "risk policy",
long_memory: "research summaries",
working_set: "current market state",
tools: "fresh data probes",
private_docs: "not public web root"
}
agent_rule = "reason only after routing"AUTHORITY SEPARATION
No single component gets to be a little tyrant.
GUARDRAILS
Default-safe until deliberately proven otherwise.
Paper trading by default
Live trading is off. Research, observation, simulation, and review come first.
No secrets in images
API keys, broker credentials, Telegram tokens, trade logs, SSH keys, and private memory never get baked into containers.
Only execution gets broker access
Researcher, trader, risk governor, maintainer, and watchers do not receive broker execution credentials.
Hard blocks on unsafe proposals
Missing stop, stale data, exceeded loss limits, active kill switch, invalid schema, or uncertain broker state means block.
Human-visible alerts
Autonomous mode still reports proposals, blocks, executions, exits, errors, and kill-switch events to the operator.
Explicit live acknowledgment
Live autonomous mode requires a specific risk acknowledgment string and current-session authorization. Enthusiasm is not permission.
human_in_loop:
trader: proposes
risk_governor: validates
human: approves
execution_bot: executes
autonomous_research:
trader: proposes
risk_governor: validates
execution_bot: paper-executes
human: alerted + kill switchTWO VARIANTS
Human-in-loop first. Autonomous research only inside a cage.
The human-in-loop design keeps approval in the execution path. The autonomous variant removes the approval gate only for controlled research and paper execution, while preserving alerts, audit logs, risk policy, and an armed kill switch.
ROADMAP
Build the skeleton before chasing the signal.
- Compose skeleton: separate services, example configs, schemas, Redis queue, and Postgres audit log.
- Paper broker: deterministic order simulation with brackets, exits, fills, and failure modes.
- Risk governor: validate every proposal against policy before any order request exists.
- Observability: healthchecks, alerts, audit viewer, daily summaries, and kill-switch verification.
- Research loop: collect outcomes by setup, session, instrument, regime, slippage, and rule-following rate.