Scan X around the clock. Build a thesis on a coin from the very first time anyone writes about it, and keep updating it until the story dies. Work out where momentum actually is, and eventually place the trade without a human in the loop.
Everything below is that note turned into something buildable, with the parts that are hard stated as hard rather than hidden.
Not one agent doing everything. Nine narrow jobs, each cheap to run and each independently checkable, because a wrong answer in one of them has to be visible instead of blended into a confident summary.
Pulls posts from a watchlist of accounts in batches, plus keyword and cashtag sweeps for tickers nobody on the list has mentioned yet. Extracts ticker, contract address, chain and whether the post is a call, a mention or paid promo.
The moment a token is seen for the first time, it gets a file. Every later mention is appended with who said it, when, and how big their audience is. The file is the coin's biography: first sighting, who picked it up, when the tone changed, when it died. Nothing is ever overwritten.
Backtests every account on its own past calls. Median return, hit rate, and the only number that really matters: how early they are relative to everyone else on the same coin. An account that is loud but late is worse than useless, it is a sell signal.
Scoring accounts answers who is early. It does not answer why anyone would buy at all. This one
reads the tweet text already stored in every dossier - for free, no extra scraping - and asks three
things: is there a case being made or just a ticker and a contract address, are people writing it in
their own words or pasting the same line, and what is the subject. On CYBERLEEK the answer came out
as gta leaks watermarked clips: a real off-chain
event people can repeat to someone who does not trade.
It has a veto. A coin with no story cannot be a buy no matter how good the timing looks, because there is no mechanism to produce the next buyer - and in a ponzi the next buyer is the entire trade. That verdict is NO NARRATIVE.
What it cannot do is say whether a story is any good. A word counter has never seen a film, a console launch or a scandal. That is agent 8.
A language model reads the actual tweets and rates how well the story recruits someone who is not already staring at a chart: what the hook is, who outside crypto could care, whether the claim can be checked against anything real, how original it is, how much of it reads as paid, and what would kill it. It runs on the Claude Code install already on this machine, so it costs nothing against the scraping budget.
Blind, on purpose. It never sees the price, the market cap, the peak or any other agent's verdict. Show a model the outcome and it will hand the outcome straight back to you as insight.
And it has no authority yet. Every judgement is stamped with the market cap at the moment it was written and never edited, so the record can convict it. Until the high-conviction calls visibly beat the low ones over about thirty judgements, the judge is printed in the dossier and changes nothing. An unscored model opinion is confident noise.
Before you ever see an alert: pool age, liquidity, is the LP burned, what the dev wallet holds, top ten holder concentration, and whether the contract can be minted or frozen. Kills most candidates outright. This is the filter that stops the fleet becoming a rugpull delivery service.
The CYBERLEEK pattern: not one caller, but more and more accounts talking about the same thing while the gaps between mentions get shorter. Fires when two or more scored accounts converge on one contract inside a window and the rate is accelerating - not on any single tweet.
And it fires again on the way out. The same counter that says get in says get out: the moment the known-late accounts and the automated whale feeds start posting the ticker, the crowd has arrived and there is nobody left to sell to. On CYBERLEEK that was whalewatchalert and dexvralisting, half an hour behind the wave. Exit alert, not a buy alert.
Running, on paper. Same code path a live wallet would take, with the signing step replaced by a ledger entry - real pools, real prices, real liquidity, pretend money. It models slippage as a function of size against pool depth rather than a flat percentage, denominates the book in SOL so a token that doubles while SOL doubles correctly makes nothing, and fills at the price at the moment the cycle runs instead of the price the thesis was written at.
There is no key on this machine. Creating one is the single irreversible step in this whole plan, and the function that would do it throws unconditionally. Turning it into working code is a human edit, so that promotion shows up in a diff rather than in a log line at four in the morning.
Holds the target, the equity curve, the vault balance and the risk envelope. Nothing else in the fleet is allowed to decide how big a position is. Size is a share of the current bankroll, so compounding needs no rule of its own.
Falling behind tightens it. Never loosens it. The instinct when a target is slipping is to size up and catch it, and that is the most reliable way to end a month at zero. The schedule is a scoreboard here, not an input to position size. The only thing that moves the envelope is drawdown, and it moves in one direction.
Build the watchlist and backtest it. No list exists yet, so the fleet seeds its own: work backwards from CYBERLEEK, CATE and EGG and find who mentioned each of them first, then second, then third. Those accounts become candidates, and their own history gets scored the same way. Deliberately slow - batched scrapes, spread over days, so nothing gets hammered. Output: a ranked list of who is genuinely early.
Every token the scouts see gets a file from first sighting. Runs read-only for a few weeks so we can look back and ask: at what point in this file was it obvious? That question cannot be answered on historical data alone, because dead tokens lose their pool history and quietly drop out of any backtest, flattering the result.
Convergence plus safety check, pushed to WhatsApp within minutes with liquidity, market cap, pool age and the thesis so far. Every alert is logged with entry price at alert time, so the scoreboard on this page is not curated afterwards.
Sized by the Treasurer, exits by rule, no discretion. It only buys what the fleet rates BUY ZONE on the cycle it runs - it will not backfill an entry into a run that already happened, which is why the table is empty today rather than broken. Compared against the benchmark below. If paper does not clear it over a meaningful number of trades, the project stops here, and that result gets published on this page too.
Two addresses. TRADE holds the working budget and nothing else, and its key would live on this machine. VAULT is generated by Armando on his own device; this machine only ever learns the public address, and half of every realised gain moves one way into it with no code path back. Whatever key sits on a machine is a key that machine can spend, so the bound is not trust, it is how little is in the wallet.
One caller was backtested properly before any of this was built. Everything the fleet produces gets measured against it - if a whole agent fleet cannot beat one guy's timeline, it is not worth running.
| Strategy | Return |
|---|
Full call-by-call breakdown: stitch-calls-concept.jouwidealewebsite.nl
Whether an AI can tell a good story from a bad one is not something to argue about. It is something to score. Every blind judgement below is stamped with the market cap at the moment it was written, so the right-hand column is what the coin did after the opinion existed.
Position size is a share of the current bankroll, so compounding needs no rule of its own. The ladder below it only ever moves down. Falling behind the goal tightens this envelope and never loosens it - sizing up to catch a slipping target is the most reliable way to end a month at zero, so the schedule is a scoreboard here, not an input.
Read live from the billing API, not from a tally the system keeps about itself. The loop checks it before every single scrape and refuses to start one that would cross the line - at the ceiling it stops scraping and keeps publishing, it does not slow down and carry on.
Measured at 0.00038 dollar per post, and only ever batched - one request covers the whole batch of ten accounts. Polling accounts separately is billed per query and would run into hundreds a day, so the fleet never does it. Two requests per batch: one window old enough to score, one small recent window that only feeds the mention ledger.
One batch of ten every 45 minutes, one at a time, with a wait between every price lookup. A batch spends 12 to 25 minutes on lookups and the rest of the cycle asleep, which clears a 48 handle queue in under four hours. Nothing here is time-critical, and the same machine also runs the CRM and WhatsApp - those come first. The NFT watchers are stopped.
A single node process under PM2, asleep almost all the time. Price history comes from a free endpoint at one request every 2.5 seconds, well under its limit. Stop it any time with pm2 stop crypto-agents.
| Date | Change |
|---|