Trading Agents Can See 8 Stock Markets. They Cannot See Tokenized Assets.

In April 2026, a data science lab at the University of Hong Kong published Vibe-Trading, an open-source trading agent. Five months later it has more than 33,000 stars on GitHub and a contributor base shipping over 150 merged pull requests per release cycle. You type a question in plain language; the agent pulls market data, writes strategy code, backtests it, and reports back. With your explicit mandate, it can also trade through your broker.

This is not an isolated project. It is the clearest example of a new distribution layer in finance: agents that aggregate brokers and data sources, then expose them to any AI assistant through a standard protocol (MCP). The user talks to the assistant; the assistant reads markets through whichever sources the stack has onboarded.

Here is what caught our attention. Vibe-Trading connects 14 brokers and 27 market data sources: US, Hong Kong, China A-shares, UK, Canada, India, Korea, crypto exchanges, forex. The number of tokenized real-world asset sources among them is zero. No RWA venue, no tokenized security dataset, no on-chain registry of real assets. The fastest-growing interface to financial markets cannot see our asset class at all.

Why agents skipped RWA

Agents consume data that is structured, verifiable and read-only. Equities have that: decades of standardized feeds, tickers, filings. Crypto has it: public chains and exchange APIs. Tokenized RWA mostly does not. The data that describes a tokenized asset – what it is, who issued it, what backs it, which rights the token carries, whether the disclosure is verified – lives in PDFs, marketing pages and scattered chain explorers. An agent cannot ground a decision in that, and a well-built agent refuses to guess.

The teams building these stacks understand this discipline better than most financial institutions. Vibe-Trading's own engineering rules are instructive: every number in a report must trace back to the tool call that produced it; a data source that cannot say what it is returns an error instead of a plausible value; a verification gate that checked nothing must fail, not pass. Agents built this way will simply not touch an asset class whose records do not meet the bar.

What closes the gap

Three things make an asset class agent-readable:

1. Verified records. A machine-readable record per asset, with each data point traceable to a source – a filing, a chain transaction, an audited document. This is exactly what Stobox Intelligence is built for: verified records of tokenized assets, with more than 900 data points per record read from a released registry.
2. Read-only data contracts. A stable API or MCP source an agent stack can onboard without trust negotiations. Read-only by construction, so nothing on that path can move funds.
3. Mandate-gated execution. When execution comes later, it comes bounded: an allowlist of assets, size caps, expiry, an instant kill switch. The agent world already treats this as table stakes; regulated RWA venues should feel at home with it.

Since 2018, Stobox has supported over $305M in tokenized assets across 100+ clients. The lesson of those eight years is that tokenization succeeds where the record is clean and fails where it is marketing. Agentic finance now enforces the same rule with software: an asset without a verifiable record is invisible to the fastest-growing channel in the market.

Tokenized assets will show up in agent stacks. The open question is whose records the agents will read.