I Hired 4 AI Agents To Run My Trading Desk

Welcome back to the Miles Deutscher weekly newsletter!

Most traders don't blow up their account on a bad trade.

They blow it up on one bad day.

The first loss is fine, but the revenge trade is the problem. Then the "make it back" trade at 3x size.

So this week I built the fix. An AI trading desk, staffed like a company, where 1 agent has a single job: say no.

Every good desk has someone whose job is to say no, while most retail traders don't.

That idea runs through the whole issue. The AI desk gets a Risk Manager. Your agents get a job description that says what they're not allowed to do.

Markets got 2 hot inflation prints, and the traders who'd sized for it slept fine. Inside MHC, our analysts posted their exits before their entries, and Morin wrote up his own mistake in public.

The edge this week isn't a better entry. It's a better "no".

IN TODAY'S READ

Your AI Trading Desk, Staffed Like a Company
Onboard Your AI Like a New Hire (4 Files)
Weekly Market update
This Week Inside MHC
New From Miles

After our main breakdown, stick around for AI Skills & Prompts, this week's Markets catalysts, and the MHC Recap of setups from inside the room.

AI X TRADING: This week's AI workflow built for traders

Your AI Trading Desk, Staffed Like a Company

I now run a big chunk of my business with AI agents.

Each one has a job title, a channel, a set of skills and a memory. They hand workto each other and post a report every night.

Then it hit me. That's exactly what a trading desk is.

A real desk doesn't have 1 genius doing everything. It has an analyst who findstrades, a risk desk that can veto them, someone keeping the books, and a bosswho reviews the week.

You can build the same thing in a weekend. Here's the team.

Agent 1:The Scanner

The Scanner watches your list and flags setups that match your rules. Nothing else.

It doesn't trade. It posts an alert in #scanner with the coin, the level and the rule it matched.

If you built the research desk from last week's issue, that's your Scanner. You'vealready hired it.

Agent 2:The Risk Manager (the one that matters)

This is the agent most people skip, and it's the one I'd build first.

Before any trade, you send it the setup. It checks 3 things: size, stop, and how much you've lost today.

If you've hit your daily loss limit, the answer is "desk closed". No debate, no "just one more".

You won't hold yourself to a rule at 2 am. An agent with no ego will.

Copy this into its role file:

You are my Risk Manager. You are not here to find trades. You are here to stop bad ones.

My rules:
- Max risk per trade: 1% of account
- Daily loss limit: 3% of account. Once hit, no new trades until tomorrow, 00:00my time
- Max leverage: 5x
- Every trade needs a stop before entry

Before every trade, I send you: the coin, entry, stop, size, and today's PnL so far.
Reply with APPROVED or REJECTED and 1 line why.
If I argue after a REJECTED, remind me of this rule: one bad day blows accounts, not one bad trade.
Never suggest a bigger size to "make it back".

Swap my numbers for yours. The number matters less than having one.

Agent 3:The Journal

Every trade gets logged: the PnL, and the reason you took it. The reason is thepart that matters.

Every night it posts a short report in #journal: trades taken, rules followed, rulesbroken, and how you felt.

That last bit sounds soft. It's where all the useful data is.

Agent 4:The Coach

Once a week, the Coach reads the whole journal and finds the pattern you can'tsee.

"You lose 70% of your trades after 22:00." "Your stops are fine, your size after aloss isn't."

Then it updates each agent's memory file, so the desk gets better every weekwithout you rebuilding it.

How to wire it up

You don't need to code. Claude Code writes the plumbing for you.

Make 1 channel per agent in Discord or Slack. #scanner, #risk, #journal,#coach.
Create a bot for each and copy its token. Treat it like an ID badge: keep itsecret.
Ask Claude Code to build a small backend that connects the bot to your AImodel and posts replies back to the channel.
Give each agent its 4 files (next section shows you how).
Deploy it, @ the bot, and test it with a fake trade.

Use your strongest model for the Coach and the Risk Manager. Judgement is thejob. The Scanner can run on a cheaper one.

Where this goes wrong

Never give an agent withdrawal-enabled keys. Read-only, or no exchangeaccess at all. You click the button.
A nightly report isn't a backtest. It tells you what you did, not whether thestrategy works.
Agents see the data. They don't see the room. They won't tell you that 3 goodtraders just cut the same coin. People still do that better.

Do this before next Tuesday (under 30 minutes)

Pick your daily loss limit. Write it in dollars.
Open a new Claude Project called "Risk Manager" and paste in the promptabove.
Before every trade this week, send it the setup. Do what it says, even oncewhen you don't want to.

That last part is the whole test.

Unlock the full agent setup (free) →

AI SKILLS & PROMPTS: Manually curated prompts, skills, workflows & more

Onboard Your AI Like a New Hire (4 Files)

Here's the mistake I see every week.

People drop an AI agent into their workflow and expect it to just work. You'd neverdo that with a new employee.

A new hire gets a company intro, a job description, training and feedback. Soshould your AI.

That's 4 files. Put them in a folder per agent, and point Claude at the folder.

→ 1. context.md: the company intro

Who you are, how you trade, what you care about. Every agent reads this one.

Fastest way to write it: let Claude interview you. Hit the voice button and just talk.

Interview me to write my context.md. Ask 1 question at a time, 12 max. Cover: what I trade, my timeframe, my account size range, my risk rules, my best and worst habits, and what "a good week" looks like for me. Then write context.md in short bullet points. No fluff.

→ 2. role.md: the job description

One per agent. What it does, what it never does, and what "done" looks like.

# Role: Journal. You log every trade I send you. For each trade record: coin, side, entry, stop, exit, size, PnL, my reason, my mood (1 to 5).

Every night at 23:00, post a report: trades, rules followed, rules broken. You never give trade ideas. You never judge. You just keep honest books.

The "never" lines matter as much as the job. An agent with no limits will try to help with everything.

→ 3. SKILL.md: the training

A skill is a process the agent can repeat. Claude has a built-in skill creator, so youdon't have to write it from scratch.

The trick: record your screen while you do the task once, and narrate as you go. Then hand it over.

Here's a recording of me doing my morning market check, with my narration. Turn it into a SKILL.md: the trigger, the exact steps in order, the sources I check, and the output format I want. Flag any step where I made a judgment call so you can ask me next time.

→ 4. memory.md: the feedback loop

This is the file that makes the agent better every week instead of the same forever.

After every task, ask me 1 question: "What should I keep, and what should I change?"Add my answer to memory.md as a short rule with today's date. Before every task, read memory.md first. If 2 rules conflict, ask me which one wins and delete the loser.

Why it works: context tells it who you are, the role tells it what to do, the skill tellsit how, and memory stops it repeating mistakes. Same as onboarding a person.

The catch: memory files drift.

After a month you'll have 40 rules, half of them stale, and the agent getsconfused. Every Sunday, open memory.md and delete anything you wouldn't sayout loud today.

Feedback is the step everyone skips. It's also the only one that compounds.

MARKETS: The catalysts and moves that matter this week

Weekly Market update

2 hot inflation prints. Oil above $100. A Korean chip crash. BTC under $80K for a few hours on Thursday.

And yet BTC only lost about 1% to 2% on the week, while Schwab, CME, and DTCC kept building rails.

My read: price is a macro trade until the Fed on Oct 28. The adoption headlines are real, but they're not a bid this week. Don't confuse the two.

→ How last week's calls aged

"Below $82,000:$80.3K, then $75K." ✅ It broke $82K on CPI day and wicked to $79,676, then reclaimed $82K. $75K never came into play.
"Above $85,500:$87K back in play." ❌ Never got there. The weekly high was $84,262.
"BTC is the cleaner hold than ETH." ✅ ETH fell about 4% to 7% vs BTC's 1%to 3%.
"Cut leverage before CPI." ✅ About $1B got liquidated on Thursday, mostly longs.

3 out of 4. I'll take it.

→ CPI and PPI: both hot

CPI came in hot: 3.7% y/y vs 3.6% expected, and core 0.3% m/m vs 0.2%.Gasoline did more than half the work.

BTC dropped about $1,100 in 5 minutes. HOOD fell 8.9%, MSTR 6%, COIN 5.3%.The 10-year touched 5.33%.

Then PPI was worse. +1.3% m/m, the biggest monthly jump since March 2022.BTC made its weekly low about 90 minutes later.

Here's the alpha. After 2 hot prints, October hike odds went down. They jumped to 44% to 49% after CPI, then faded back to about 28% after PPI. Chair Warsh gave no hint in Bangkok: "I don't prejudge meetings."

The market thinks the Fed doesn't want to hike. So a hold is mostly priced, and ahike on Oct 28 would be a real surprise.

→ Key takeaways from this crypto week

1. MSCI is kicking "non-operating" companies out of its indexes, and Strategy is on notice.

MSCI approved the new rules on Thursday. The final list lands Nov 11 and takes effect Nov 30. MSTR fell 9.6% on Friday.

Why it matters for your bags: Strategy only bought 250 BTC last week. A lot ofSaylor's buying now runs through STRC, so the STRC vote on Oct 28 matters more than the headline. If STRC holds $100, the buying can restart.

2. Schwab opened crypto to every eligible client.

That's more than $13T in client assets. BTC and ETH are live now, with SOL, AVAX, and LINK coming by year-end.

Why it matters for your bags: distribution is the slowest bid and the stickiest. The SOL, AVAX and LINK listings are a year-end catalyst most people aren't tracking.

3. CME listed UNI and BCH futures today.

The contracts go live Oct 19. UNI is already down 30%+ from its $10.9 peak.

Why it matters for your bags: a regulated venue means institutions can hedge and short. Listings are often "sell the news" first. Watch first-week open interest before you get excited.

4. The tokenized-stock fight started.

9 companies objected to their shares trading on OKXICE, including Broadcom, Verizon, and CoreWeave. Meanwhile, DTCC set its launch scope: 25 securities, onCanton and a private network.

Why it matters for your bags: last week I said don't chase the headline. Tokenization is happening on permissioned rails first. The public-chain winners come later.

5. The $91M Ledger lesson: the threat was in the box.

Tampered devices from a Southeast Asian reseller read the 24-word phrase during setup and sent it out over cellular. About $91M from 474 addresses. Ledger's own chip wasn't compromised.

→The catalyst calendar

Wed Oct 21, ~17:50 EEST: Solana Alpenglow goes live. Finality drops from about 12.8s to about 150ms.
Wed Oct 21, after the US close: Tesla, Meta and IBM report. AI stocks have dragged crypto all month.
Thu Oct 22, 03:00 EEST: Aerodrome and Velodrome become Aero. It isn'tautomatic; holders must upgrade via the portal.
Fri Oct 23 to Sat Oct 24: TON unlocks $49M, and Humanity (H) unlocks 7.3%of supply.
Wed Oct 28, 20:00 EET: FOMC decision, presser at 20:30. Strategy's STRCvote is the same day.

MHC RECAP: Setups and wins from inside Miles High Club

This Week Inside MHC

These are positions and updates our analysts shared in the room between Oct 13and Oct 19, not recommendations. Levels and reasoning are theirs.

Last week's scorecard:

Fabian on $AERO: took the rest of his profits just under 0.90, from a 0.804entry, over a week before the launch. He may rebuy a sell-the-news dip.
Miles on $BAT: still running when I posted my update, and I'm paying myself on the way up. I've left some on for the news.
Paradise on $ZEC: sold at 1300. ZEC rejected hard right after.
Surf Club on $BTC: stopped out. Wicked out of the rotational play, then the D1swing. Small losses, as planned.

3 green, 1 red, and the red one was sized to be boring.

→ Miles on $ZK ( ▼ 2.21% ).

Vitalik just said the zkEVM speed race is won and the focus should move to security. Last time he put attention on ZK, the token moved hard.

There's also a ZKsync announcement teased for Oct 21 on institutional assetsonchain. I'd rather get a retest of the money noodle first.

The big IF is BTC. If it loses $80K, I cut fast.

→ Fabian on $ZEC ( ▲ 2.44% )

Fabian started DCA'ing back into ZEC at 1130, adding back what he sold last month.

He sees the 1100s as value, just above the 1k psychological support. Below 1k on a market-wide crash, he'd "buy that with both hands".

→ Surfxbt on $ETH ( ▼ 0.47% ) .

A high-volume sweep of the range lows took out every stacked low built up last month, right into the weekly trend.

Surf calls it a higher-timeframe rotational trigger that doesn't come around often.He went long with 1/3 size before PPI, with bids resting lower for the rest.