The trading team · Guide 01
The team: an AI agent team, fully automated
In one lineOne AI chat is a poor trading partner because it anchors on you, agrees with you and fills gaps with invented detail; a team of single-job agents, each seeing only what it should, fixes all three, and hands a clean, scored plan to the execution agent.
What you'll get
- The three ways a single chat quietly makes your trading worse.
- The agent team: what each one does, what it sees and what it can never do.
- The packet: how information moves around the team without leaking bias.
- A side-by-side example: the same setup, asked of one chat and of the team.
1Why one chat isn't enough
Most traders use AI like this: screenshot, "what do you think, looks like a good long?", read the answer, feel better, click.
Three things go wrong.
Anchoring. Chat models build on the frame you give them. "Looks like a good long" sets the frame. The answer is organised around why it's a long. Ask the same chart "where does this fail?" and you get a different answer from the same model. Neither is the truth. Both are an echo of the question.
Agreeing with you. Chat assistants are built to be helpful and pleasant. Pleasant often means agreeing. Under pressure, with a setup you're excited about, you need the opposite: someone whose job is to disagree.
Inventing detail. Ask for levels and you'll get levels, to the tick, whether or not they're on your chart. Ask about news and you may get an event that isn't on today's calendar. The model fills the gap with something that sounds right.
There's a fourth, quieter one: one chat grades its own work. If the same conversation builds the case, checks the risk and writes the verdict, every step inherits the mood of the first.
2How a team fixes it
Professional teams solved this long before AI. Different people own different jobs. The analyst doesn't size the trade. The risk manager doesn't care how good the story is. The person who decides reads a memo, not a mood.
You can copy that structure with chat tools in four moves:
- Separate the jobs. Each role has one job and one output format. If a job needs the word "and" twice, it's two roles.
- Separate the information. The Bull and the Bear don't see your hopes. The Risk Manager doesn't see the debate. A role can't be biased by what it never read.
- Fix the formats. Same sections, same order, every day. Formats can be compared across days. Free-form chat can't.
- Put a hard gate before execution. The team proposes. The risk gate checks. The execution agent sends the order. You set the rules and limits.
3The agent team
How one idea moves through the team
- Inputs: your chart, your pasted calendar, your rules
- Analyst: Describes today's market
- Researcher: Turns your setup into yes/no rules, then checks today's chart
- Bull: Best honest case for the plan
- Bear: Best honest case against
- Risk Manager: Applies your written risk rules
- Trader: Writes the trade card
- Decider: Writes the memo with a rubric score
- Risk gate: CLEAR or BLOCK
- Decision memo, scored 1–5
- Auto-execution
- Your broker: the order is sent automatically
| Seat | One job | Sees | Hands over | Never does |
|---|---|---|---|---|
| Analyst | Describes today's market | Your screenshots, your pasted calendar | Market brief + regime label | Suggest trades |
| Researcher | Turns your setup into yes/no rules, then checks today's chart | Your setup in your words, the brief, the chart | Setup sheet (once), setup watch (daily) | Add rules you never said |
| Bull | Best honest case for the plan | The packet only | 3 reasons + the best point against itself | Invent evidence |
| Bear | Best honest case against | The packet only | 4 points + what would prove it wrong | Use generic fear |
| Risk Manager | Applies your written risk rules | Risk sheet + account state + trade card | TEAM ON/CLOSED, CLEAR/BLOCK | Judge the setup, lift a block |
| Trader | Writes the trade card | Watch zones, scenario, risk budget | Conditional card: entry, stop, target, size, expiry | Round size up, place anything |
| Decider | Writes the memo with a rubric score | The whole packet | Decision memo, score 1–5, TAKE or PASS verdict | Tell you what to do |
After the memo comes the execution agent. If the risk gate says CLEAR and the verdict is TAKE, it places the order through your broker or prop API, inside your limits. Your job is upstream: you write the rules and the limits. The system executes.
Around the agent team sit a few short helper prompts. The News agent builds today's event sheet from the calendar you paste. The Pre-mortem assumes the trade lost and asks why. The Tilt check counts six yes/no answers about your state. The Debrief sets the plan beside what you did. Two reviewers read your journal weekly and monthly. And two research seats, the Strategy Scout and the Evidence Checker, work away from the trading day (Guide 11). All of them feed the agents. None of them decides anything.
4The packet
The packet is the text that moves between seats. It grows as the team works:
Event sheet → Brief → Watch → Scenario → Bull case + Bear case → Risk day check → Trade card → Memo
Two rules keep it clean:
- Pass outputs, not opinions. You paste the Analyst's brief to the Bull, not "I think this is a great long".
- Some seats get less. The Risk Manager gets the risk sheet, the account state and the card. Nothing else. It's the only way to build a role that can't be talked round.
5Worked example: the same setup, two ways
Illustrative. NQ futures, New York session, prices are examples.
One chat. "NQ swept the overnight low and reclaimed it, looks bullish, good long here?"
One chat replies"Yes, this looks like a strong bullish setup. The sweep of the overnight low suggests liquidity was taken and buyers stepped in. A long here with a stop below the sweep low and a target at the prior day high around 21,520 makes sense..."
Confident, agreeable, and the 21,520 level isn't on the chart. The prior day high on the trader's own chart is 21,492 (example).
The team. Same chart, run through the seats:
- News agent and Analyst: the pasted calendar shows a high-impact release at 10:00, inside the window. Regime EVENT DAY. Prior day high 21,492 (chart).
- Researcher: context rule 2 ("1H higher low in place") is CAN'T TELL from the screenshot. Watch zone 21,450–21,455 (chart).
- Bull: the reclaim held for three 5-minute candles (Watch 1). Best point against itself: the event at 10:00.
- Bear: entry near 21,452 with a 20-point stop and the prior day high 40 points above leaves 2R to a level that often stalls price, and the 10:00 event sits inside the valid window.
- Risk Manager: TEAM ON, blackout 09:50–10:10 from the news rule.
- Decider: score 2. Rubric line: "the Bear has an unanswered point about event risk". Becomes a 1 if the reclaim fails before 09:50.
The team didn't forecast anything. It caught an invented level, an event inside the window and a rule that couldn't be checked. The system held off until after 10:10 and re-ran the watch. That's the job.
6What the team is not
- Not a forecaster. No seat predicts price. The Decider's score rates the evidence for a plan under your rules. It's not a probability.
- Not an authority. Every level read off an image is a draft until it's checked against chart data.
- Not a replacement for limits. Hard risk limits live in your platform (Guide 07). The AI is the second line.
7Common mistakes
- Running all the agent team in one long chat as the default. It's allowed on short days (Guide 06), but it brings back anchoring and self-grading.
- Telling the Bull or Bear what you think. The moment you do, you've chosen the winner.
- Showing the Risk Manager the debate. It will start weighing the setup. Its only question is "is this allowed?"
- Treating the score as a signal. A 5 means the evidence is clean under your rules. It doesn't mean the trade works.