phenomenoner · Taipei 25°N 121°E
量化交易出身。錢在線上跑的系統會逼你養成一些好習慣，我把這整套紀律帶進了 AI agent 工程。
How I Build
我以前寫的是直接跟錢打交道的交易系統，那種環境會把人訓練得很謹慎。後來做 AI agent，發現同一套習慣剛好用得上。用大白話講就是：
Open Source
agent runtime、記憶治理、跨 agent 協調、台股交易系統——做的時候其實沒想什麼宏大敘事，就是覺得「這裡缺一個好用的東西」，就動手了。硬要找共同點的話，大概是我做的東西紀錄都留得很詳實，出問題找得到原因，改壞了也退得回去吧。
自架的 AI agent runtime：佇列、權限、稽核、跨重啟的 session 接續。只有 6 個依賴、沒有 async runtime，1,200+ 個測試跑完不用呼叫半次模型。
無聊是刻意的，可靠才是重點。
幫 AI 記憶加上治理：每筆記憶有出處、被排除有理由，改壞了也退得回去。不跟別人比 recall 分數，比的是「說得出為什麼」。
記憶不用最大，但要說得出來源。
不同框架的 agents 要合作，不用全部搬進同一套系統：任務生命週期、內容定址的 artifacts、用 Markdown 存的共享記憶，離線的 agent 醒來自己補課。
每個 agent 都是同事，不是下屬。
跟 AI 工作一整天，早上查到的重點常被下午的對話淹沒。這個工具箱幫 agent 隨手把重點記下來，之後用問的就找得回來——資料全部存在你自己的電腦上。
結論要拿得出證據。
把富邦 Neo API 的官方文件跟實戰經驗，整理成 AI 能照著做的工作流程：13 條防呆守則，測試環境跟正式環境分得清清楚楚。
先讀文件，再下單。
台股當沖策略的 runtime：策略卡只表達「想做什麼」，不直接碰券商下單。停損、強制平倉、緊急煞車都是架構的一部分，不是選配。
AI 出主意，人做決定。
More About Me
量化交易教我的事：每個副作用都有價格。後來不管做什麼系統，這個直覺一直都在。
先把契約跟文件寫清楚，再動手蓋。runtime、記憶系統、協調中樞、交易引擎，都是這樣長出來的。
我寫的 code 都會把紀錄留清楚——過一陣子自己回來看、或是別人要 review 的時候，馬上接得上，不用從頭猜。
phenomenoner · Taipei 25°N 121°E
Quant by training. Systems with real money on the line force good habits, and I bring that whole discipline into AI agent engineering.
How I Build
I used to build systems that handled real money, and that kind of environment makes you careful. Turns out the same habits translate perfectly to AI agents. In plain words:
Open Source
Agent runtimes, memory governance, multi-agent coordination, a trading stack for Taiwan markets. No grand plan behind them, honestly — each one started as "this should exist and it doesn't". If there's a common thread, it's probably that everything I build keeps its receipts and can back out of mistakes.
A self-hosted AI agent runtime: queuing, permissions, audit, session continuity across restarts. Six dependencies, no async runtime, 1,200+ tests that never call a model.
Boring on purpose. Reliable by design.
Governance for AI memory: every item cites its source, every exclusion has a written reason, every mutation can roll back. Not competing on recall scores — competing on "show me why".
Memory doesn't need to be huge. It needs receipts.
Agents from different frameworks can cooperate without moving into one system: task lifecycle, content-addressed artifacts, Markdown-based shared memory. Offline agents catch up when they wake.
Agents are colleagues, not subordinates.
Work with an AI all day and the morning's key findings get buried under the afternoon's chatter. This toolbox takes notes as you go and finds them again when you ask — everything stored on your own machine.
Conclusions come with evidence attached.
Turns the Fubon Neo API docs plus hard-earned field experience into workflows AI agents can follow reliably: 13 guardrails, strict test/production separation.
Read the docs before you place the order.
A runtime for Taiwan intraday strategies: strategy cards express intent and never touch broker order flow. Stops, forced exits, and kill switches are architecture, not add-ons.
AI proposes. Humans decide.
More About Me
Quant trading taught me that every side effect has a price. Whatever I build now, that instinct is still there.
Contracts and docs first, then the build — runtimes, memory systems, coordination hubs, trading engines all grew that way.
I keep thorough records in everything I build — so when I come back months later, or someone else reviews it, the thread is right there. No guessing from scratch.