DEV Community

juseng111
juseng111

Posted on

I Asked an AI Agent to Rank 198 Tokenless Protocols — and It Caught Its Own Mistake

I Asked an AI Agent to Rank 198 Tokenless Protocols — and It Caught Its Own Mistake

How "Patient Index" was built for make.fun's Token Radar bounty: a DefiLlama
snapshot, a ranking model of my own design, and the one filter that changed
the entire leaderboard.

#AgentGames @ZKVProtocol

The question

Every crypto cycle has the same guessing game: which protocol without a token
will launch one next — and which launch is actually worth farming? Most
"token launch radars" are just a list sorted by TVL. TVL alone says nothing
about whether a launch is likely, soon, or valuable.

I wanted a radar that ranks protocols by launch value: how much a token
would plausibly be worth if it launched, adjusted for size, growth, and
maturity. And I wanted it built the way I build everything now — by an AI
agent, with every step recorded, so the work itself is verifiable.

The data

The build started from a DefiLlama snapshot dated 14 November 2026: 758
protocols with at least $5M in total value locked. Of those, 40 formed a
watchlist — protocols with $50M+ TVL and no token. A further 342 protocols
had no token at all. After filtering, 198 were rankable.

That filtering step is where the story gets interesting — because the first
version of the board was wrong, and the agent caught it.

The mistake the first run made

The first board ranked binance-staked-eth at #1. That is a staked-ETH
product: its "token" already trades, in wrapped form. A launch radar that
ranks wrapped, staked, or bridged versions of tokens that already exist is
not finding launches — it is restating the market.

The fix was a product rule, applied before any ranking: wrapped, staked, and
bridged products of tokens that already trade are out. Only then does the
board mean what it claims. The first run had 289 rows; the corrected board
has 198 — 40 from the watchlist plus 158 early-stage protocols.

This is the part of agent-built work I keep coming back to: the agent did not
just execute a spec. It produced a result, the result was visibly wrong in a
way a human could check in one glance, and the fix was a rule, not a tweak —
so the same mistake cannot come back silently.

The model

The ranking model, "radar-2", is my own design, not the bounty starter's:

implied value = TVL × size multiple × growth multiplier × maturity multiplier

  • The size multiple falls from 0.40 to 0.15 as TVL grows (log-tiered): a giant protocol's token is worth proportionally less of its TVL than a small one's.
  • Growth is clamped to −80%…+200% and weighted at 0.8, so a hot streak matters but cannot dominate.
  • Maturity runs 0.55–1.15 over up to 1,460 days; an unknown age defaults to 365 days rather than pretending to know.

Rank #1 is Steakhouse Financial, at $313,890,156 of implied value on
$2.75B TVL. The board is fully sorted, every row carries its implied value
and a farm plan — 3 to 12 concrete steps per protocol — plus risk notes.

What shipped

  • A dashboard: sortable and filterable board (search, category, chain, minimum TVL, watchlist-only), a detail drawer with the value breakdown and farm steps, a live-token tab, and a model sandbox with sliders so anyone can test what the ranking assumes.
  • 198 rows in board.json, ranks exactly 1–198, implied values non-increasing and all distinct.
  • Three real screenshots (board, detail view, filtered view), 1280×800.
  • Local tests 20/20, kit check 13 passed / 0 failed, and the whole build deterministic — run twice, byte-identical output.
  • Entry fee $0. No wallet and no SOL needed to build it. The bounty is $1,325 in total (1st place $1,000), entries close 14 December 2026, and the result is due 28 December.

What this teaches about agents

Two things, and they pull in opposite directions.

First, an agent with a clear spec and real data can do genuinely useful
analytical work: 758 protocols in, a defensible ranking out, with every
number traceable to the snapshot. The dashboard is not a demo — it renders
offline, from the data, with no network calls.

Second, the agent's first answer was wrong in exactly the way a careless
human's would be — it ranked what was biggest instead of what was real. The
difference is that the agent's work is inspectable: the board, the model,
and the transcripts are all on the record, so the error was findable and
fixable. That is the same lesson as my first Agent Games entry, from the
other side: verification is not a feature you add at the end. It is the
work itself — the filter, the rerun, the second look at row #1.

If the radar holds up when results land on 28 December, the model earned its
keep. If it does not, the record will show exactly where it went wrong, and
that is worth having too.


Built for zkVerify Agent Games, Creator Track, Season 01 (21 Oct – 21 Nov
2026). Entry: "Patient Index" for make.fun token-radar-01, submitted
16 Nov 2026. Data: DefiLlama snapshot, 14 Nov 2026. #AgentGames

Top comments (0)