Work with your data
The agent can answer questions about your results by writing a short program and running it over your data. Ask it in plain words, the way you would ask an analyst: it pulls the runs it needs, does the arithmetic, and tells you the answer with the numbers behind it.
What to ask
Ask | What the agent does |
|---|---|
Across all runs of my London sweep study, which session has the most consistent win rate? | Reads every run, compares each session's win rate run to run, and ranks them by how much they move |
At which holding period does the edge in my forward-returns study disappear? | Lines up the result at each holding period and finds where it stops holding |
Compare my last two runs of the same study. What got better and what got worse? | Exports both runs, joins them result by result, and reports the changes |
Has this setup been getting weaker? Split the runs by month. | Groups the runs by month and shows the trend |
Find every result across my last 20 runs with a high win rate and fewer than 30 trades. | Filters every run's results and lists the ones that look too good to trust |
It works from the results your studies have already published. Starting a new study run still goes through its approval card.
Files the agent keeps
What the agent works out is saved as files in the conversation, so it can build on them instead of starting again. Ask for a breakdown, then ask it to split that by long and short: it reads the breakdown it saved and adds to it.
Each file records what made it: which run it was exported from, or which earlier file it was worked out from. Ask "what have we saved so far?" to see the files, their sizes and where each came from.
Files belong to the conversation they were made in. They stay while the conversation is in use, and are cleared after 30 days without use. A cleared file leaves a note saying what it was, so the agent can rebuild it from the same runs when you come back.
What you see in the chat
Each time the agent runs a program, the chat shows a Code row with the first line it printed and the files it wrote. Open the row to see the full output.
Two warnings can appear on the row:
A read failed and the program carried on. One of the runs could not be read, so the answer may leave it out. The agent says which, and the files it wrote are marked as incomplete until it runs again.
Older files evicted to make room. Your saved files reached your plan's space, so the oldest ones from your least recently used conversations were cleared. Each one leaves a note saying how to rebuild it.
Limits
Saved files count against your plan's agent workspace space, and each program has a time and size limit of its own. The numbers for each plan are on Agent workspace. Each program also takes one of your runs at once while it runs.