AI

What an AI Agent Actually Needs to Score a Token End to End

An AI agent scoring a memecoin needs price, liquidity, holder, social, and wallet data — not vibes. Here's the honest stack.

· 7 min read · Blackhat Empire

An AI Agent Is Only As Good As Its Feeds

Everyone is shipping an "AI agent" that scores memecoins. Most of them are a language model with a price feed duct-taped to a Twitter scraper. That is not a scoring engine, that is a horoscope with a chart.

If you are building one, or paying attention to one, you need to know what data it actually has to touch to produce a number that means anything. Here is the full pipeline, layer by layer, and where each layer lies to you.

Layer 1: Identity and Contract Truth

Before any scoring happens, the agent needs to know what the token is. That means contract address, chain, deployer address, creation timestamp, mint authority status, freeze authority status, and whether the contract is verified or proxied.

On Solana this is mostly about mint and freeze authority plus token program quirks. On EVM chains it is ownership, proxy patterns, and hidden mint functions. An agent that scores a token without checking whether the deployer can still mint unlimited supply is scoring a rug with a nice logo.

This layer is binary and boring. It should be able to instantly disqualify a token, not blend into a score.

Layer 2: Liquidity and Pool Structure

Price means nothing without pool depth. The agent needs LP total value, LP locked or burned status, who owns the LP, and how concentrated the pool is. A token with $40k liquidity and a single LP holder is a different animal than one with the same number spread across three pools and burned.

It also needs to know the pair's age and whether liquidity was added in one shot or dripped in. Dripped liquidity after a pump is a distribution schedule, not organic growth.

Layer 3: Holders and Distribution

Raw holder count is the most gamed metric in this market. The agent needs the top 10 and top 50 holder percentages, excluding known LP and burn addresses, plus how many of those wallets were funded from the same source.

Cluster analysis is the difference between "1,400 holders" and "1,400 wallets controlled by four people." If your agent cannot cluster funders, it cannot tell you anything real about distribution.

Layer 4: Volume and Trade Flow

Total volume is a vanity number. The agent needs buy versus sell volume, unique trader count, average trade size, and wash-trade detection — repetitive size patterns, circular wallets, and same-block round trips.

It also needs to separate fresh buys from recycled volume. A token doing $2M in volume with 90% of it from 12 wallets is not trending, it is being farmed.

Layer 5: Wallet Quality and Smart Money

This is where most agents are weakest. The agent needs a wallet reputation layer: which wallets have historically entered early on tokens that survived, which ones exit into strength, and which ones are serial deployer-adjacent.

Without this, "smart money bought" is just "a wallet bought." You can track this kind of flow yourself through the alert channels listed at blackhat.finance/channels.html, or join the chain-specific groups like BH GMGN SOLANA and BH GMGN BASE to see how different feeds frame the same event.

Layer 6: Social and Attention Data

Social data is noisy but not useless. The agent needs mention velocity, unique account ratio, account age distribution, and whether the same handful of accounts are posting the same copy across every token.

An honest social score penalizes coordination. A dishonest one rewards it because coordinated shilling looks like "high engagement."

Layer 7: The Scoring Function Itself

The hard part is not the data, it is the weighting. Every layer above can be gamed individually. The scoring function has to assume that and punish correlation — if volume, social, and holder growth all spike from the same wallet cluster in the same ten minutes, that is one signal, not three.

A useful agent outputs a score and the reasons behind it, so you can see which layer drove the number. If it cannot explain itself, it is not analysis, it is a black box with a countdown timer.

Where You Fit In

You do not need to build this to use it. You need to know what your tools are missing so you can stop treating a score as a verdict. Cross-check the metrics that matter against our metrics reference, watch how events get framed across alerts, and hold yourself to basic rules before you size anything.

Most memecoins go to zero. No agent changes that. The best an agent can do is help you avoid the obvious traps faster. When you want to check a chart or a flow for yourself, do it on GMGN or the mirror at GMGN — and read the numbers, not the narrative.

Community

Stay connected across the chains:

Charts and on-chain research: https://gmgn.uk.