OpenClaw Trader is less interesting to me as a trading bot and more interesting as an autonomy problem. I wanted to see what happens when the primary user of the system is not a person watching a dashboard, but an agent reading structured output and deciding what to investigate next.
That changes the design. Everything important has to emit JSON. The desk has to keep its own state, decisions, audits, reviews, lane metrics, and memory in files the agent can read. Nothing critical should live only in chat history. If the agent cannot reconstruct what happened, the system is not really autonomous. It is just vibes with a cron job.
The desk is split into lanes: momentum options, mean reversion, trend continuation, equity breakout, overnight swing, and a random-timing baseline. Each lane can have its own autonomy state and risk profile. New lanes start in observe-only paper mode. That matters because most ideas should not be allowed to place orders just because they compiled.
The constraint is risk and evidence. The repo has levers, but they are bounded. Stop loss, take profit, position size, trailing stop behavior, max positions, cash reserve, and drawdown limits can be adjusted only inside policy. Per-lane overrides matter because one lane getting better does not mean the whole desk should loosen up.
The most useful pattern is the agent brief. The agent starts by reading a daily, weekly, monthly, or all-time brief: performance, lane deployments, no-trade reasons, desk health, open proposals, investigation questions, and lever signals. That turns the system into something inspectable instead of something that just says it is working.
Notes for next time: autonomy needs boring contracts. Structured outputs, fail-closed behavior, paper-first operation, clear lane lifecycle, audit logs, and human-gated live mode are not extras. They are the difference between an experiment and a system I can reason about.
Build log