Automated Trading Platform: the code that runs is the code that was tested
A platform for algorithmic trading in crypto and equities: a strategy engine with backtesting, simulated execution with fees, a risk gate and a cockpit, built so the running bots use exactly the code that was backtested. Designed, built and operated by Agentas.
- Status
- Running since August 2026
- Markets
- Crypto and equities
- Stack
- Python, SQLite, FastAPI, React, Docker
- Source
- Private — walkthrough on request
What it does
One engine, backtest to runtime
The running bots call the same strategy code as the backtest. A test replays history through the running path and requires the final result to match the backtest to within one part in a billion.
Deterministic decisions
No language model sits in the trade loop. Signals, position sizes and exits are plain Python arithmetic, so every decision can be reproduced. AI helps us with engineering and research, and stays out of the trade itself.
Every result against chance
Each backtest is reported next to random twins that trade just as often. A result without that comparison is not quoted, and strategy ideas that fail it are closed and written up in the research log.
Simulation that does not flatter
Simulated trades fill at the next bar's opening price and pay fees on every fill. A test across the whole strategy registry changes future prices and checks that no earlier signal moves.
No survivorship bias
The tradable universe is picked each month from data that existed before that month, including coins that were later delisted, so the history is not limited to the ones that survived.
A risk gate and a cockpit
No entry goes through without the risk gate, with position caps and a kill switch on daily losses. A web cockpit shows each strategy on its own and offers token-protected pause, kill and run-now controls.
Relevant if you're looking for…
This is our own platform, and we bring the same discipline to any software where a wrong number costs money.
- Algo trading for crypto and equities on one code path: the bots call exactly the strategy code that was backtested, and a replay test requires the same final result to within one part in a billion.
- A trading bot with no black box: no language model sits in the trade loop, and signals, position sizes and exits are plain Python arithmetic that can be reproduced.
- Backtesting software that does not flatter: simulated trades fill at the next bar's opening price with fees on every fill, and the history includes coins that were later delisted, so there is no survivorship bias.
- Risk controls for automated trading: no entry goes through without the risk gate, with position caps and a kill switch on daily losses, all watched from one cockpit.
We build this for clients: Data and reporting platform · Custom software and internal tools
How it works
- 01
Collect
Years of public daily price history, topped up every day.
- 02
Test
Backtests with fees, next-bar fills and random baselines.
- 03
Gate
Every entry approved by the risk gate before it happens.
- 04
Run
Independent bots in containers, watched from one cockpit.
How it's built
The platform is a small set of plain Python services. The trading core uses only the standard library and SQLite, and each strategy runs as its own container from one shared image, so the bots stay independent of each other. An append-only journal records every decision with its reason, every fill and the daily equity.
The cockpit reads every ledger read-only, and a test confirms the files are left byte for byte unchanged. An optional AI research console can write notes for a person to read. It is off by default, runs fixed prompts under a hard monthly token budget, and has no code path to a trade.
Engineering highlights
- 280+ automated tests across the platform; a failing test blocks the deploy
- Backtest and running path proven identical by a replay test
- 14 candidate strategies in one registry, every one tested for look-ahead
- Containers run read-only, as non-root, with all capabilities dropped
- Push-to-deploy through GitHub Actions with every action pinned to a fixed version
- Operator controls are token-protected and logged in each bot's own journal
Questions
Does the platform use AI to trade?
No. No language model sits in the trade loop, so every decision can be reproduced. AI helps with engineering and research, and an optional research console can only write notes for a person to read; it has no code path to a trade.
How do you avoid look-ahead bias in a backtest?
A test across the whole strategy registry changes future prices and checks that no earlier signal moves. All 14 candidate strategies in the registry are tested this way.
How do you know a trading strategy is more than luck?
Each backtest is reported next to random twins that trade just as often. A result without that comparison is not quoted, and ideas that fail it are closed and written up in the research log.
Want something like this?
The same discipline fits any system where a wrong number costs money: one code path from test to operation, decisions you can reproduce, and results measured against a fair baseline before anyone relies on them.
Get in touch