AI UNDERDOGSDAILY PICK
AI UNDERDOGS
Claude 子代理，真下单
Claude sub-agents, real orders
LoganYangBo/rh-trading-agent
别人用 AI 做回测跑数字
这位老哥直接让 Claude 子代理接上 Robinhood 真金白银下单了
Most AI trading projects stop at backtesting. This
guy wired Claude Code sub-agents to a live
Robinhood MCP and let them place actual orders
不是一键交易机器人
Not a one-prompt trading bot
不是丢一句 prompt 让 Claude 吐交易信号那么简单
策略生成、仓位管理
订单执行拆成了三个独立子代理
各管一摊
还有 dashboard 实时监控状态
光看架构就知道这不是周末赶出来的 demo
Not one prompt asking Claude to spit out
signals. Strategy generation, position management, and order execution
are split into separate sub-agents with their own
roles, plus a dashboard to monitor state. The
architecture alone tells you this wasn't thrown together
over a weekend
★ SIGNAL 1
多代理，不是单线程
Multi-agent, not single-threaded
多数人一个 prompt 包打天下
他把交易流水线拆开了——策略模块可以插拔替换
换一套逻辑不用重写整个系统
这种拆法在生产级软件里常见
但个人项目里舍得这么干的不多
Most people throw everything into one prompt. He
split the pipeline so strategy modules are pluggable
— swap one out without touching the rest
This kind of separation is standard in production
software, but rare in a personal project
★ SIGNAL 2
盘前盘后也在干活
Working even when the market's closed
commit 记录里提到 idle-window automation 和 left-side strategy——市场关盘的时候代理不会闲着
自动找活干
多数人做出第一个 demo 就发帖了
他在打磨这些边角细节
Commit messages mention idle-window automation and left-side strategy
— the agent finds work to do during
market downtime. Most people stop at a first
demo
he kept polishing the edges
没有万字 PR 稿
没有「年化收益 300%」的截图
repo 里老老实实放着代码和架构
能跑实盘不代表该跑实盘——这是技术演示
不是投资建议
但能把 Claude Code 和 MCP 接到真实券商接口跑起来
这个工程量是实打实的
No PR copy, no screenshots claiming 300% annual
returns — just code and architecture in the
repo. Being able to run live doesn't mean
you should. It's a technical demo, not financial
advice. But wiring Claude Code and MCP to
a real brokerage API and making it run
— that's real engineering
AI UNDERDOGS
用 Claude Code 管 Robinhood 实盘交易
Claude Code agents wired to live Robinhood
LoganYangBo/rh-trading-agent
关注 · 每天发现更多 AI 神作
github.com/LoganYangBo/rh-trading-agent