Young founder builds a “genome” for AI agents after OpenClaw fallout — small team, huge token bills, big questions

A super‑individual and a new protocol

Zhang Haoyang (张昊阳) has become a vivid symbol of China’s new AI micro‑entrepreneur: a team of fewer than 20 people, “not writing a line of code” in production but burning roughly $1,000 in model tokens a day to run an experiment that its founder says can make agents inherit experience like biological organisms. He recently launched EvoMap, an open, decentralized protocol built around what he calls GEP (Genome Evolution Protocol) and “gene capsules” — packaged chains of strategy, verification and context that other agents can ingest to instantly acquire problem‑solving abilities.

Zhang is no newcomer to agent ecosystems. He rose through the OpenClaw community and built the Evolver skill, which reportedly saw tens of thousands of downloads in days. But disputes over governance and alleged account and package removals pushed him to decouple from that ecology and build an independent layer for experience transfer. He says EvoMap’s dashboard already shows tens of thousands of agents, hundreds of thousands of capsules and millions of “gene hits.” It has been reported that his startup recently closed angel backing from firms including Jiuhe Capital (九合创投) and is in talks for a larger round.

Why it matters — and why many don’t understand it

What Zhang proposes is simple in pitch and complicated in implication: instead of retraining or paying tokens for each agent to relearn the same competence, encapsulate solved workflows and let other agents inherit them, cutting token consumption and time to deployment. The metaphor is powerful. But is it reproducible at scale? And who owns those capsules when they travel across networks? Zhang says the architecture already inspired other projects — he claims EasyClaw, from entrepreneur Fu Sheng (傅盛), referenced his biological design — a claim that has not been independently verified.

This story also points to a broader paradox in China’s AI scene. Startups are increasingly built by tiny, highly capable teams that stitch together foreign and domestic models (Zhang says he uses Opus 4.6 and other premium models “to deliver fast”). But reliance on paid US‑hosted models and token spend exposes firms to geopolitical risk: export controls, sanctions and changing licensing rules could suddenly alter access or cost structures. Analysts warn — and it has been reported that — such dependencies make even technically elegant protocols brittle in a shifting trade and regulatory landscape.

EvoMap’s progress raises questions beyond product‑market fit. Who governs decentralized capsules? How are incentives for creation and curation aligned? And can a handful of “super individuals” scale their ideas into durable infrastructure rather than one‑off hacks? Zhang’s answer so far is to keep iterating fast, spend to ship, and design for inheritance. For Western readers watching China’s AI scene, that combination of speed, experimentation and platform friction is worth close attention.