My Grok memecoin desk: 6 agents, $100 to $4,352 in two weeks
two of them find things: wo weeks, around 75 trades, 68% win rate, memes across multiple chains. none of that came from me staring at charts at 3am. it came from six agents splitting the work, and me doing the one part that can't be delegated: the signature.
this is not a backtest and not a "what if". this was a live run, and below is how the desk is wired, step by step, in the order I'd build it again from zero.
I don't hunt tickers anymore. six seats, one thread, one report every evening. on most days my whole job is to read that report and type APPROVE on the trades that need me.
01. start with a manager, not a trader
the first instinct is to open grok and ask it to find you a 10x. wrong agent, wrong seat. my first agent is HEAD OF DESK, and it has never placed a trade. it turns my goal into daily tasks, keeps every token moving through the same pipeline, and brings me only the decisions that need a human.
before I trusted it with anything, I gave it a test I could check in under a minute: list the ten most traded new tokens on one chain over the last hour, with contract address and source for each, then double-check two of the addresses itself. if it gets that wrong, it doesn't get anywhere near a wallet.
02. the wallet is a burner, not a bank
the run started with $100 on purpose. a fresh wallet that holds nothing else, so the worst case is a number I chose before the first trade, not one I discover after. memes rug, agents misread, chains congest. the burner is how you survive all three.
one rule sits above everything else on the desk:
// non-negotiable
no seed phrase or private key ever goes into a chat. not once, not "just to test".
no agent connects a wallet. I connect it, I approve it, every single time.
if an agent asks for a key, the run is over and the agent is deleted.agents read, scan and prepare. they never hold keys. that's the whole security model, and it's the first thing I'd teach anyone copying this.
03. job descriptions, not prompts
a prompt is a one-off question. a job description is a role that survives the next session. every agent on my desk gets three blocks: what it owns, what good work looks like (written as checks, not adjectives), and where it stops and asks me.
the third block is what makes this work. without a hard stop line you're not running a desk, you're gambling with extra steps.
here's HEAD OF DESK, exactly as I'd write it today:
You are my Head of Desk.
// you own
Turn my weekly goal into daily work for the desk.
Run every token through: found -> checked -> confirmed -> staged -> closed.
Give each token a lead ID. Every message about it carries that ID.
Morning status and an evening report, every day.
// good looks like
I hear from you twice a day, plus once per decision that needs me.
Nothing reaches SNIPER without a RISK pass and a WHALE confirm.
I can read any trade start to finish by its lead ID.
// you stop
You never trade, never sign, never touch keys.
You never skip a stage because a token "looks obvious".
You never change a risk limit. Only I do, in writing.
If the day's loss hits my daily limit, you halt the desk and tell me.04. six specialists, not one genius
one agent doing discovery, contract checks and execution ends up average at all three. worse, when a trade goes wrong you can't tell which part failed. split by job, not by task size.
two of them find things:
// SEARCH - scout
Own: new pairs and fresh launches on the chains I trade, dev wallet deploys,
the Telegram channels and X accounts I gave you.
Good: a short list of leads each day, every lead with an on-chain reason.
Stop: you never recommend an entry. Leads go to RISK, nobody else.
// WHALE - smart money
Own: the wallet list I gave you. Early buys, clusters, wallets that usually move first.
Good: you spot accumulation before the chart shows it.
Stop: you confirm or you stay quiet. You never size a position.// RISK - contract and rug check
Own: mint and freeze authority, LP lock, top holders, dev wallet history,
honeypot and tax checks, on every chain we touch.
Good: most leads die here, and every kill has a written reason.
Stop: you never make an exception. One hard fail = dead lead.two of them have hands:
// SNIPER - execution
Own: prepare the order once a lead is checked and confirmed.
Size, entry, stop, targets, slippage, fee.
Good: the order is staged before I finish reading the message.
Stop: nothing gets signed without my literal "APPROVE #id".
Not a summary. Not "looks good". The exact words.
// EXIT - position manager
Own: every open position. Trailing stops, taking profit in parts, invalidation.
Good: profit is taken on the way up, not given back on the way down.
Stop: you never add to a position and you never average down.05. teach it one real trade, including a "no"
describing the checklist in a paragraph is slow and vague. so I walked one trade start to finish inside the chat: here's the lead, here's what I check, here's why I'm in. then a second one, where I killed the token halfway and said exactly which check failed.
if every example you give ends in a buy, the desk learns what buying looks like and nothing about passing. and passing is most of the job. on memes, by far.
06. one thread, one goal for the week
every agent reports into one thread, so the whole life of any trade can be read in one place by its lead ID. I don't hand the desk a to-do list. I hand it a goal for the week: the target, max risk per trade, max daily loss before everything stops and waits for me.
a to-do list means I've already done the thinking. a goal lets HEAD break it down, which is the whole point of having more than one agent.
07. the line is the signature
decide what needs you, or the agents will decide for you, and they'll get it wrong. my rule: anything reversible, they do alone and log it. anything that signs a transaction, moves money or touches a risk limit stops and waits for me.
one exception: if RISK flags a rug on an open position, getting out is pre-approved. on memes, the speed of the exit is the difference between a small loss and a zero.
08. automate the boring parts
RISK and EXIT check every open position on a schedule and only speak when something changed. silence means nothing moved, and that silence is worth a lot: it means I don't have to look.
the evening report request, word for word:
Every day at 21:00, send me the desk's evening report:
found by SEARCH, killed by RISK, confirmed by WHALE, executed by SNIPER.
Open positions, PnL today, win rate, best and worst token.
One thing that worked, one thing to change tomorrow.
If any source is down, say so and continue with the rest.
09. read the numbers honestly
75 trades at 68% means roughly 24 losers. the desk didn't win by being right every time. it won because EXIT cut losers fast and RISK killed the worst tokens before anyone could buy them. I stopped counting the rugs it caught.
and the honest part: two weeks of memes in a hot market is a run, not a track record. $100 to $4,352 is what happened, not what you should expect. I kept the desk. I didn't keep the expectation.
what this actually gives me
hundreds of tokens scanned, most of them killed before I ever saw them, and a handful of trades that reached me already checked, confirmed and staged. I'm not the one digging through new pairs, reading contracts and babysitting exits anymore. six agents split that, and I read one report a day.
the setup is simpler than it looks. open grok, build HEAD OF DESK first. give the other five strict job descriptions, the way you'd onboard a new hire. walk them through one real trade and one rejected one. one thread, one lead ID per token, one evening report. no VPS, no code, no developer.
a memecoin desk used to mean no sleep, forty open tabs and a wallet you were scared to look at. now it's six agents, one thread, and one word from me: APPROVE.
if you're going to build one, save this first. you'll want the job descriptions open next to you while you write your own.
I'm posting how this desk evolves from here: what the agents get wrong, the rugs RISK catches, and the numbers, wins and losses both. follow @A_Pirate_I if you want the next run.
and if you've already built a desk of your own, tell me which agent you'd hire as number seven.