The coding agent
An assistant that writes, fixes and runs your strategy, with @file mentions.
The Talios agent is a coding assistant that works in your project: it writes and edits your files, checks its own code, runs the backtest and reads the results with you.
Setting it up
The agent uses the same model connection as the AI analyst: your own key from Anthropic (Claude), OpenAI (ChatGPT), Google (Gemini) or OpenRouter, picked in the model picker of its message box. See Connect a model. Its conversations are separate from the journal assistant's and saved on your account. Keep the tab open while it works: closing it stops the answer.
Your journal
The agent can read your Afterburner journal, only when you ask: "look at my EUR/USD trade of March 3", "how did my London trades do?", "compare this backtest with my journal". It never reads it on its own, and never changes it.
What to ask
| Request | What it does |
|---|---|
| "Write an FVG strategy: retests with the EMA 200 trend, London and NY killzones" | Creates the files, checks them, runs the backtest and sums up the results. |
| "Add a news filter: no trade 30 min around high-impact news" | Edits your strategy with near_news() and runs it again. |
| "Which pair and session carry the edge?" | Reads the run's trades and answers with numbers. |
| "Run the backtest and explain the results" | Runs, then walks you through what matters. |
| "Move the stop to break-even at 1R" | Changes the code where needed and shows you the diff. |
| "Backtest it on GBP/USD 15m over 2023" | Changes the pairs, timeframe or period, then runs. |
How it works in your project
- You see every change. Each step shows as a row: files listed, read, created, edited (with the diff), code checked, backtest run (with PF and net R). Its writes appear live in the editor, and the files it touched keep a green mark.
- It finishes clean. After writing code it checks it and fixes errors, and ends on a clean check and a successful run, or tells you what blocks it after three attempts.
- It asks first before deleting or overwriting files you wrote yourself, and asks back with options when your request is ambiguous.
- It never touches your money settings. It can change pairs, timeframe and period, but account size and risk stay yours, in the Backtest panel, and the timezone is the project's for good.
- It knows what you changed. Files you edited by hand since its last answer are pointed out to it, and so is a project switch.
- While it works, the editor and the file tree wait (Agent editing…), and Run is paused until it is done.
Pointing at a file with @
Type @ in the message box to list your project's files, keep typing to filter, then Enter or Tab to pick. The file becomes a blue tag in your message, and its current content is sent with your question: "why does @pullback.py never return True on Mondays?".
Numbers, not opinions
The agent only reports figures from the backtest it ran, formats them in R, and calls a result on fewer than about 30 trades anecdotal. It answers in your language.