Skip to content
leve.devAI Agents·Case Study 2 / 7
AI Agents·2025

pnyx.vc

Traditional VC: six-month decision cycles, closed networks, relationship investing. Pnyx flipped it — anyone could submit a token thesis, five named AI agents judged it on merit, and profitable calls earned their submitters 30 %.

Multi-Agent AIn8nSolanaSupabaseNext.js
5
named AI agents
30 %
profit share to submitters
0.1–0.5
SOL per position
100K
views on launch thread
111
commits before launch
Pnyx position history: open positions with realized and unrealized PnL, plus the live trade log
The public position history — every trade visible, winners and losers alike.
01 — The Idea

VC decisions in minutes, judged on merit.

The launch thread said it plainly: traditional VCs are broken — six-month decision cycles, closed networks, human bias, relationship investing over merit. Meanwhile the best opportunities move at light speed.

Pnyx was the counter-model: an autonomous on-chain fund where the community sourced the deal flow. Anyone could submit an investment thesis for a Solana token. A panel of five AI agents — each with a name, a role and a specialty — evaluated it and decided. Every position was public.

02 — How It Worked

Five agents, one decision.

A submitted thesis ran through the full panel before a single lamport moved:

  • Athena evaluated the submitter's Twitter history — credibility, context quality, shill signals.
  • Archimedes verified the submitter's past token calls on-chain, measuring each from tweet timestamp to peak.
  • Pnyx, the decision maker, weighed both inputs and sized positions between 0.1 and 0.5 SOL.
  • Sokrates monitored the portfolio and flagged exits; Eirene explained everything to users in chat, with live portfolio access.
  • When a call was profitable, 30 % flowed back to the person who submitted it.
One of the n8n analysis workflows behind the Pnyx agents, roughly 40 nodes from webhook to structured output
One of the analysis workflows behind the agents — three parallel research branches (tweet, author, community) feeding a structured-output model.
03 — The v1.1 Sneak Peek

Chart structure meets caller credibility.

A later iteration, tested on a dedicated wallet, pushed the analysis further: it combined chart structure — break of structure, Fibonacci pockets, fair-value gaps, momentum across timeframes — with a credibility score for the human making the call, and let that score modulate position size. The test wallet is no longer active; the output below shows how it reasoned.

Token chart with the Pnyx test wallet's buy and sell markers, average buy at 390K and average sell at 754K market cap
The v1.1 test wallet on the chart — average buy at 390K market cap, average sells at 754K.

The five agents

Each agent had a name, a personality and a single job on the panel.

AthenaTwitter history evaluator — credibility, context quality and shill signals of the submitter
ArchimedesCall performance evaluator — past token calls, measured on-chain from tweet to peak
PnyxDecision maker — weighed all inputs and sized every position
SokratesPortfolio evaluator — monitored positions and portfolio health
EireneChatbot — explained the platform and answered with live portfolio data
“Trader: CryptoBro987 scores 59.5 — moderately credible but not hot — so I'm giving a small size bump, not an aggressive one.”
— Pnyx v1.1 analysis output

Planning something similar?

I design and ship AI agent systems, data platforms and full-stack products — from first idea to production.

Get in touch