> For the complete documentation index, see [llms.txt](https://documentation.justbeep.it/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://documentation.justbeep.it/product-overview/riffs/agent-infrastructure.md).

# Agent infrastructure

Agents run off-chain on Google Cloud and use the latest US frontier models in a mixture-of-models decision pipeline. No single model can place a trade on its own.

<figure><img src="https://2348763252-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Ffrgr0Kse7wRCOlaPSZd7%2Fuploads%2F3LfW9ZrXiSaerSGHKiBf%2Friffs-agent-loop.png?alt=media&amp;token=18855f43-4f70-4e4a-b97f-3891838b9398" alt=""><figcaption></figcaption></figure>

\
Signals feed a proposer model, other models critique its trade, and the order goes ahead only if they agree and the guardrails pass. Every executed trade carries a plain-English note for the user.

* **Riff compiler.** Parses the view into a thesis, maps it to the tradable asset universe, scores each candidate on thesis fit and liquidity, and sizes the basket with per-name caps.
* **Mixture-of-models pipeline.** Each decision passes through several models with distinct roles: one proposes, others critique and stress-test, and the pipeline acts only when they converge. Disagreement means no trade.
* **Signals.** Fundamentals, technicals, market data (price, depth, funding rates) and news, refreshed continuously so the agent reacts to markets that never close.
* **Deterministic guardrails.** Models propose; a rules layer outside the models decides what can execute. It enforces allowed assets and venues, order size, leverage and the user's limits, and the wallet's own policies enforce them again at signing.
* **Explanations.** Every trade carries a plain-English note on why it happened, shown to the user next to the trade.