·HOOX·2 min read

[The Watcher That Never Sleeps: agent-worker]

agent-worker: cron every 5 minutes, reconcile D1 and exchanges, close positions or hit the kill switch.

hooxagent-workerrisk-managementcronaikill-switchautonomous-systemstrading

Sources: docs/devops/workers/agent-worker.mdx, signal lifecycle stage 7, data-flow AI risk sequences, paper limitations on cron risk.

Thesis

Webhooks are reactive. Markets are continuous. HOOX’s answer to “who manages risk when no alert fires?” is agent-worker: a cron-driven isolate that wakes every five minutes, reads truth from D1 and the exchanges, and is empowered to close positions and shut the front door.

Execution (trade-worker) must stay fast and boring. Policy (agent-worker) must stay periodic, auditable, and allowed to slam the kill switch. Merging them makes every prompt tweak a deploy to order-signing.

1. Charter

Concern Behavior
Trigger Cron */5 * * * *
State Open positions via d1-worker
Market Public mark prices from CEX APIs
Actions Service Bindings → trade-worker
Alerts telegram-worker
Config Trailing stop %, take-profit %, drawdown, AI provider chain

This is stage 7 of the signal lifecycle—asynchronous, deterministic first, LLM second.

2. Normal loop

Cron → agent → D1 OPEN positions → exchange marks
     → compare KV watermarks / drawdown
     → optional CLOSE / scale-out via trade-worker
     → telegram notify

Trailing stops store high-water marks in CONFIG_KV.
Take-profit scale-out secures gains without a second TradingView alert.
Portfolio hygiene as code—not as a human staring at Telegram at 03:00.

3. Kill switch — hard stop

When drawdown breaches configured limits, the agent writes trade:kill_switch = true (CONFIG_KV). Consequences:

  1. hoox refuses new trades (503 path on ingress pre-flight).
  2. Flattening commands to trade-worker.
  3. Emergency Telegram alert.

Gateway and agent share a single source of truth for “is the system allowed to open risk?” That shared key is more important than any model temperature.

4. Multi-provider AI

  • Primary: Workers AI
  • Fallbacks: OpenAI, Anthropic, Google AI, Azure OpenAI
  • Health-checked failover, streaming, vision/reasoning where configured
  • Periodic NL health digests over system_logs
  • Vectorize RAG collaboration with telegram-worker (/ask, search)

Runtime config (GET/POST /agent/config) mutates provider chain without redeploying the mesh. Invariant: models summarize and advise; they do not replace signed risk gates on the placement path.

5. Academic honesty

The paper is careful: the cron risk manager is intentionally simple. Event-driven risk on ExchangeConnectionManager (fills, partials) is planned—and bulked up in Enterprise (hibernatable WS + Workflows). Marketing a 5-minute loop as HFT risk would be academic malpractice.

Previous: Latency race · Next: Multi-tier storage

← All postsHome