Own the sovereign-agent stack, fade agent-token hype

Thesis

Agents are learning to act, acquire compute and copy themselves. What stops them is money they can hold without KYC and compute they can rent without an account. This basket goes long the parts of the agent stack that agents demonstrably use, and short the agent-token narrative whose usage is self-reported and subsidized.

Why now

• Capability is here. UK AISI RepliBench: models "excel" at provisioning compute and handling crypto but "fail completely at our Know Your Customer check." Palisade (May 2026): an open 27B model on a single A100 hacks its way onto new hosts and copies itself in 33% of runs.
• The payment rails are being built ahead of demand. The x402 Foundation (Visa, Mastercard, Stripe, Google, AWS, Cloudflare) has formed, and Chainalysis counts 138M+ x402 payments on Base.
• But usage is still small and in USDC. 99.6% of x402 value settles in USDC, real 30-day volume on Base is about $1.1M, and TRM attributes only 0.6–7.5% of it to agents. So own the layers with working mechanisms, not the tokens claiming an "agent economy."

Legs

Pearl (PRL): a 10% option, not a core holding

Pearl pitches "money for machines… without KYC": the one asset an agent could mint by computing. That makes it the only liquid bet on the money slot of the stack. It is sized as an option because the product doesn't match yet:

• Pearl's inference API takes cards only.
• The team calls its L1 "not a good fit for agentic payments."
• ≤0.27% of recent hashrate even matches the shape of an AI model layer.
• Supply grows ~96% over the next 12 months.
• io:PRL longs pay ~0.043%/h funding (~370% APR, ~1%/day). The board won't show that drag.

Scale the sleeve up if PRL-denominated inference or an x402 facilitator ships. Cut it if funding stays above ~0.02%/h or PRL loses the ~$1.5–2 consumer-GPU breakeven.

What confirms / kills it

Confirms: agent-attributed share of x402 rising (TRM/Chainalysis); Venice or NEAR AI publishing agent usage numbers; a stablecoin freeze, frontier-API restriction or agent-wallet crackdown, which reprices permissionless rails as an escape valve.
Kills: agent payments stay under ~$5M/month and USDC-only (the thesis collapses to COIN); regulation targets KYC-free agent compute; NEAR/VVV give back their 90-day runs (+185% / +128%). Hard stop at basket −20%.

The board tracks price only. Funding is not included: about 11% APR on the alt perps, paid by longs and received on the VIRTUAL short, plus ~370% APR on the PRL sleeve (≈3% of the basket per month at 10%).