> For the complete documentation index, see [llms.txt](https://docs.usance.xyz/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.usance.xyz/use-it/sentinels.md).

# Sentinels

A **Sentinel** is a bounded autonomous agent that acts on your account within a mandate you sign, for example repaying a little debt to hold a safety buffer when the account approaches its limit. It is the mandate model made autonomous, with the boundaries kept explicit.

The one rule everything else follows: **AI is not a financial authority.** A Sentinel may *observe* state and *propose* an action, but a deterministic policy and your signed [mandate](/use-it/mandates.md) *decide* what is allowed. Every action is checked as:

```
AllowedAction = ProtocolAllows ∧ MandateAllows
```

re-read against live state immediately before it submits.

## What a Sentinel can never do

Even a fully compromised agent key:

* cannot exceed the mandate you signed
* cannot borrow, trade, or open a hedge
* **cannot withdraw your collateral** (there is no such action to delegate)
* cannot widen its own permissions

You can **pause or revoke** any Sentinel at any time. Revocation is immediate and permanent, and a revoked mandate refuses the agent's next attempt on-chain.

## The library

The **Sentinel Library** lists published strategy templates from the committed on-chain manifest, showing publisher, version, risk class, the exact actions and trigger classes it requires, its fee model, and audit status. A template holds **no authority and contains no executable code**: it is a declarative, versioned manifest. Statistics are shown as zero until real runs exist, never as invented ROI.

### Safety Buffer (T1)

The first template. It is **risk-reducing only**, so it may `REPAY` and `ADD_COLLATERAL` and nothing else, and it acts to keep your account above a safety threshold you configure.

## Arming one

From **Arm a Sentinel** you see the full permission preview, every action it may take and the explicit list of what it cannot (borrow · trade · withdraw collateral), **before your wallet opens**. Arming is two transactions: the EIP-712 **mandate** the agent acts under, then the **instance** that pins the template and its manifest hash. A mismatched template is refused, never guessed.

## The autonomy plane

Every run passes through the same auditable sequence, and each step writes evidence:

**observe → trigger → snapshot → plan → validate → authorize → reserve → execute → reconcile → receipt.**

A run that is refused before submission is a real, recorded outcome. The receipt says so rather than implying an action occurred. A public run timeline at `/sentinels/runs/…` shows exactly what a Sentinel observed, why it acted, what it executed on-chain, and what changed, with no wallet required. A live autonomous repay is recorded on the [proof explorer](/understand-it/networks.md).

Publishers can read the publishing contract at `/developers/sentinels`. See [For developers](/build-on-it/developers.md).


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.usance.xyz/use-it/sentinels.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.