AI

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

A serious agent doesn't guess. It pulls contract, liquidity, holder, flow, and social data in a fixed order — here's the stack and where it breaks.

· 7 min read · Blackhat Empire

The Agent Is Only As Good As Its Inputs

Everyone wants to bolt an LLM onto a token and call it an "AI agent." Most of what passes for that is a chatbot reading a price ticker. A real scoring agent does something harder: it collects a defined set of data, in a defined order, and outputs a score with a reason attached. If the data layer is thin, the score is astrology with a terminal font.

So let's walk the pipeline the way an agent should. This is education, not a buy signal. Assume every token here can go to zero, because most do.

Stage 1: Contract and Provenance Data

Before anything else, the agent needs to know what it's looking at.

  • Contract address and chain — the single source of truth. Everything else keys off this.
  • Mint/freeze authority status on Solana, ownership and proxy status on EVM chains.
  • Deployer wallet and deploy timestamp — who launched it and when.
  • Token creation route — launchpad, bonding curve, direct deploy, or migrated.

This stage is binary. If the contract has an active mint authority or a hidden owner function, the agent should stop scoring and flag it. No amount of volume fixes a token that can be printed into oblivion. The reference metrics that matter here are documented in our metrics reference.

Stage 2: Liquidity and Structure

A token with no liquidity is a screenshot, not a market.

  • Pool depth and how it's split across venues.
  • Liquidity-to-market-cap ratio — thin float relative to cap means a small sell nukes the chart.
  • LP lock or burn status, and who holds the LP.
  • Concentration of supply — top holders, insider clusters, and wallets that funded each other.

An agent that skips holder clustering will happily score a token where ten wallets are one person. That's not analysis, that's being farmed.

Stage 3: Flow and Volume Quality

The loudest number in memecoins is also the most lied about.

  • Volume across time windows — 1h, 6h, 24h — and whether it's rising or decaying.
  • Buy/sell pressure and average trade size distribution.
  • Unique wallet counts versus raw transaction counts. One bot can be a thousand trades.
  • Net smart-money and whale movement — who's accumulating, who's exiting.
  • Fresh wallet activity — new addresses funded right before a run are a tell.

The agent needs to distinguish organic churn from wash volume. If the same wallets round-trip the same size, the score should drop, not rise. Our alert stack tracks flow from multiple angles — the alerts reference lays out which channels cover what, and you can see the full directory at blackhat.finance/channels.html. Our chain groups are where traders compare notes: BH GMGN SOLANA @gmgnx_solana, BH GMGN BSC @gmgnx_bsc, BH GMGN BASE @gmgnx_base, and BH GMGN CHAT @gmgnx_chat.

Stage 4: Social and Attention Data

Attention is real data, but it's the easiest to fake.

  • Mention velocity — is chatter accelerating or already flat?
  • Account quality — follower graphs, account age, posting patterns.
  • KOL involvement — paid or organic, and whether they disclosed.
  • Narrative fit — does this token have a story, or just a ticker?

An agent should treat social as a weighting, never a trigger. A hundred bot posts is not conviction. This is where most retail gets wrecked: they buy the loudest thing, not the strongest thing.

Stage 5: Risk and Exit Conditions

The final stage is the one almost nobody builds.

  • Rug-pattern signatures — honeypot behavior, sudden authority changes, LP pulls.
  • Slippage at realistic size — what a normal position actually costs to exit.
  • Invalidation rules — the exact conditions under which the score flips to "no."

An agent without invalidation logic isn't an agent, it's a hype amplifier. Ours should be able to say "this is dead, walk away" as fast as it says anything else. The rules of engagement are in our rules reference.

Where You Actually Use the Output

A score is a filter, not an order. Use it to cut a thousand tokens down to ten worth a second look, then verify by hand on GMGN at gmgn.uk — or the gmgn.fr mirror if the main portal is slow. Cross-check the contract, the holders, and the flow yourself. The agent narrows the field; you still pull the trigger and own the outcome.

The Honest Bottom Line

Good data in a fixed order beats a clever prompt every time. Build the pipeline, keep the invalidation rules sharp, and remember the house truth: memecoins are extremely high risk, and the majority go to zero. An agent that helps you avoid bad tokens is worth more than one that helps you find the next one.

Community

Stay connected across the chains:

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