The trading team · Guide 10
Going fully automated: the go-live guide
In one lineAutomation is the last step, not the first. You're ready when your setup sheet is fully mechanical, your journal shows you follow it, and you know exactly what a fully automated team needs and what it takes to run properly.
What you'll get
- The five levels from a chat team to a fully automated one, and what changes at each.
- What a fully automated team actually needs, without the hype.
- The go-live test: six questions that tell you if you're ready.
- When to hold off, and what to fix first.
1The five levels
| Level | What runs it | What you do | What's new |
|---|---|---|---|
| 1. Chat team | Agent roles in chat (this library) | Run the routine, read the verdicts | Separate roles, written rules, a risk check |
| 2. Scheduled team | Scheduled prompts prepare parts of the routine | Read and check the outputs | Less typing. Still no chart access or order connection |
| 3. Alerts and checks | Platform alerts fire on your trigger; the check runs on each one | Review the alerts and checks | The trigger is watched for you |
| 4. Semi-automated | Software places orders from your rules, inside your limits | Watch fills and logs at small size | Rules written as code, risk limits enforced in code |
| 5. Fully automated | Code finds, sizes and places trades inside limits enforced in code | Supervise, review, switch off | Testing, monitoring, a kill switch, hosting |
What never changes, at any level: you set the rules and limits; the system executes inside them. At levels 1–3 it runs on demo. At 4 it runs live at small size. At 5 it runs fully automated: you set the rules and limits, and you can switch it off any time.
2What a fully automated team actually needs
Honestly, it's a lot more than "connect AI to my broker":
- Rules a computer can follow with zero judgement. Every CAN'T TELL in your setup sheet has to become a calculation.
- Clean data. A reliable price feed, and a way to notice when it's late, missing or wrong.
- Testing on history that's honest. Fills at realistic prices, costs included, no peeking at future bars, and testing on data the rules weren't tuned on. A pretty backtest shows very little. Guide 11 covers the ways a test fools you.
- Stress tests. Thousands of reshuffled orderings of the trade list, to see the bad runs to expect before they happen.
- A risk gate in code. The same limits as your risk sheet, enforced by code that runs after any AI and before any order. An AI can add a veto. Nothing can lift one.
- Execution. A broker or platform connection that allows automated orders, with permission from your broker or firm.
- Monitoring and a kill switch. Logs of every decision, alerts when something breaks, one control that cancels everything and stands the team down.
- Somewhere to run it. A machine or server that stays on during your sessions.
- Maintenance. Platforms change, data feeds change, firms change rules. Someone has to keep it working.
None of this is impossible. All of it takes skill, time or money. Often all three.
3The go-live test
Answer yes or no. Six yeses before you build anything:
- Mechanical. Could a stranger follow your setup sheet with no questions? Has every rule been checkable (no CAN'T TELL from the rule's wording) for 8 weeks?
- Followed. Has your journal shown
rules_followed = Yon nearly every trade for 8 weeks or more? - Logged. Do you have a long enough journal that the setup's behaviour, in R and in losing runs, is something you can describe, with a sample you've written down and an Evidence Checker report (Guide 11) whose sums you've checked?
- Bounded. Can you state, as numbers, the worst day and the worst run of losses you'd accept before switching it off?
- Allowed. Have you confirmed in writing that your broker or firm allows automated trading on your account type?
- Owned. Do you have the time and skill to build and maintain it, or the budget for someone who does?
If any answer is no, stay at level 1–3 and keep journaling until it's yes. That's not failure. It's the evidence the automated system needs.
4Reasons to hold off (for now)
- Your edge still relies on judgement you can't yet write as code. Fix the wording first.
- Your rules change often. Automation needs stable rules.
- Your account or firm doesn't allow it.
- The team is already doing its job. If the routine keeps the account out of bad trades and inside its limits, take the time to collect evidence before stepping up.
5Worked example: two traders, same question
Illustrative.
Trader A has run the team for 12 weeks. The setup sheet hasn't changed in 9 weeks. Every context rule is a calculation: time windows, prior day levels, swing comparisons on the 1H. The journal shows rules followed on all but a handful of trades, each with a note. The firm's rules page allows automation on this account type, confirmed by email. A has written down the worst run of losses that would mean switching it off. Six yeses. A starts at level 3 (alerts on the trigger), then looks at level 4.
Trader B has run the team for 5 weeks. Context rule 2 is still "the 1H looks like it's trending", which the Researcher keeps marking CAN'T TELL. The journal shows 7 rule breaks, all after losses. Two yeses. B stays at level 1, works on rule 2's wording, and makes the post-loss pause the weekly change. Automating now would only automate the breaks.
6Common mistakes
- Automating to escape discipline. Code doesn't fix a broken rule. It runs it faster.
- Automating a setup you haven't journaled. You'd be automating a story.
- Trusting a backtest without stress tests and unseen data. Most good-looking backtests fall apart on data they weren't built on.
- Skipping the broker or firm permission check. Read the rules. Get it in writing.
- Letting an AI place orders with no code-enforced limits after it. The risk gate goes in code, after the AI, every time.