Available
Research agent
Improves a strategy toward your goal with backtests, then checks the finished version on a held-out final test.
AI agents · powered by Claude
Put an AI agent to work on your trading algorithm.
Give it a task and a budget. It edits a copy of your strategy, runs backtests on the LEAN engine and tests its finished version on dates it never saw. You review the result and decide what to keep.
- Final test on unseen dates
- Pay only for what it uses
- Nothing changes until you adopt
TaskKeep drawdown under 15% with a Sharpe above 1, 2010–2020
Your versionAgent's version
- readRead main.py and the last backtest
- testBacktest 1Sharpe 0.71 · DD 31%
- editAdded a 200-day trend filter
- testBacktest 4Sharpe 1.12 · DD 18%
- editScaled position size by volatility
- testBacktest 7Sharpe 1.38 · DD 13%
- finalFinal test · 2019–2020Sharpe 1.06 · DD 14%
$1.84 of $5.007 backtestsIllustration
Works with
- Python strategies
- LEAN engine
- Claude by Anthropic
- Point-in-time datasets
- The AI assistant
- Your prepaid balance
What are AI agents?Research you'd do by hand,
Research you'd do by hand,
done in the background.
Change something, backtest, read the report, repeat. An agent runs that loop for you, on the same engine and data as your own backtests.
01
Ask
Describe the task in plain English: a lower drawdown, a higher Sharpe, fewer trades, a trend filter to test.
02
Research
On a copy of your code it makes a change, runs a backtest, reads the statistics and decides what to try next.
03
Decide
See every step, the changes as a diff and a final test on unseen dates. Adopt the result, or leave it.
The agents
What your agents can do.
One agent is ready today. Two more are on the way, built on the same safeguards.
Coming soon
Monitoring agent
Will watch your running deployments and tell you when one behaves differently from its backtest.
Coming soon
Paper trading agent
Will run a strategy on live prices without real money, and report how it compares with its backtest.
How a run works
From a task to a result you can trust.
- Give it a taskPick an algorithm, describe what you want and set a budget from your balance.
- It iteratesChange, backtest, read, repeat: one backtest at a time, every step on the record.
- One final testIts finished version runs once on dates kept back from the start.
- You adopt, or notCompare the results, read the diff, and copy the changes into your algorithm.
Final testIllustration
Tested once on 2019–2020, after 6 development backtests on 2010–2018.
| Development | Final test | |
|---|---|---|
| Sharpe ratio | 1.38 | 1.06 |
| Drawdown | 13% | 14% |
| Annual return | 11.2% | 8.9% |
| Orders | 412 | 96 |
The gap between the two is how much of the development result was luck. A strategy that only fits its tuning dates shows up here.
30%of your dates kept back for the final test
$1smallest budget, paid only as it's used
1backtest at a time, in an isolated container
0changes to your code until you adopt them
Safeguards
Built so you can trust the result.
An agent that tunes a strategy until the backtest looks good can fool you. These rules are there to stop that.
A held-out final test
The agent develops without the final test dates and runs them once at the end. The last 30% of your dates are kept back unless you pick your own.
Your code stays yours
The agent works on a copy. Nothing in your algorithm changes until you review the diff and choose to adopt it.
Spending you control
Each run has a budget you set, taken from your balance only as it's used. Stop a run at any time, or add budget to continue.
Isolated backtests
Each backtest runs in its own container with no network access, the same way your own backtests do.
Every step on the record
The run shows what the agent read, what it changed and which backtest it ran, so you can follow its reasoning.
No real-money trading
Agents research and backtest. Deploying to a broker account is always your decision.
Assistant or agent
Two ways to build with AI.
Use the assistant while you write, and an agent when there's a search to run.
AI assistant
Beside you in the editor
- Best for
- Drafting a strategy, explaining code, fixing errors
- Backtests
- Runs one when you ask
- Changes
- Applied in the editor as you go
- Overfitting check
- Your own judgment
AI agent
On its own, in the background
- Best for
- Searching for a better version: filters, sizing, parameters
- Backtests
- As many as its budget allows, one at a time
- Changes
- Kept on a copy until you adopt them
- Overfitting check
- A final test on held-out dates
Your agent's first task is a minute away.
Top up your balance, pick an algorithm, and describe what you want in plain English.
FAQ
Questions, answered.
Anything else? The datasets and algorithmic trading pages cover the rest of the platform.
What is an AI trading agent?
An AI agent is a model (Claude) that works on its own toward a goal, instead of answering one message at a time. On MarketsCoder an agent takes a goal for your algorithm, such as a lower drawdown, and works through it in the background: it changes the code, runs backtests, reads the reports and decides what to try next.
Does the agent trade with real money?
No. Agents work on a copy of your algorithm and only run backtests. Your algorithm doesn't change until you adopt the agent's version, and you decide what is deployed to a broker.
How do agents avoid overfitting?
Each run keeps back a final test period: dates the agent can't see while it develops. When it's done, its finished version is tested once on those dates, and you see both results side by side. A strategy that only works on the dates it was tuned on shows up there. Leave the dates empty and the last 30% of your algorithm's dates are kept back.
How much does an agent run cost?
You choose a budget for each run, and it has to be in your prepaid balance to start. You pay only for what the agent uses. When the budget runs out the agent pauses, and you can add more or stop it.
What can I ask an agent to do?
Anything you can measure in a backtest: reduce drawdown, raise the Sharpe ratio, trade less often, hold up across a longer period, or test an idea such as a trend filter. Clear, measurable goals give the best results.
What's the difference between the AI assistant and an AI agent?
The assistant works with you in the editor, one request at a time, and you see each change as it happens. An agent works alone in the background toward a goal, for as long as its budget allows, and hands you a finished result to review.
Which AI agents are available?
The research agent is available now. A monitoring agent for running deployments and a paper trading agent are coming; neither is available yet.
Which AI model do the agents use?
Claude, by Anthropic. It's the same model family as the AI assistant, given more steps and a budget to work through a task on its own.