# How it works

> The Layer 1 Agent, the browser extension, the backend and the console: how Holistic AI EndLayer turns device facts into graded, evidenced verdicts.

Holistic AI EndLayer, the Guardian Agent Platform, has four parts. Each has one job, and the line between them is deliberate.

| Part | What it does |
|---|---|
| **The Layer 1 Agent** | Runs on each macOS, Windows and Linux device. Collects 26 signals every 15 minutes, delivers them to your organisation's EndLayer service, and applies what your policy asks. |
| **The browser extension** | Installed by your device management tool in Chrome, Edge, Brave, Vivaldi and Chromium. Checks prompts and uploads on AI websites and hands each one to the agent on the same device, which decides. |
| **The backend** | Receives reports, grades them against your policy, keeps history for the windows you set, and runs the connections to your other systems. |
| **The console** | Shows the fleet as tables and charts: devices, people, tools, MCP servers, cost, events and evidence, plus Governance, where policy is set. |

## Facts on one side, verdicts on the other

**The agent reports facts.** It says that a coding agent is set to skip permission prompts by default, or that some AI domains on your required blocklist are reachable in a browser. It does not decide whether that is acceptable. A device is the wrong place to keep a policy: a laptop that can be told the rules can be told different rules.

**The backend grades those facts.** Each signal has its own grader, and a grader is a function of the evidence and the policy alone. That makes every verdict reproducible, and history can be graded again when a rule changes.

Three rules keep the grading honest:

- **Unknown stays unknown:** where the agent could not measure something, the result is recorded as unknown and is never graded as a pass or a fail.
- **No policy means ungraded:** if nobody has set the policy input a check needs, the result is ungraded, never a quiet pass. An empty allowlist means "nobody has decided yet".
- **Each state is counted on its own:** pass, fail, ungraded and unknown are reported separately, so a pass rate never includes results that were never measured.

**The agent then acts on the policy.** It places the prompt scanner where the policy names, removes what an administrator approved for removal, and halts what a kill switch names. See [what it stops](https://www.endlayer.ai/docs/enforcement).

## One outbound connection

The agent makes outbound HTTPS calls to **one place: your organisation's EndLayer service**. It sends no telemetry to anyone else, it does not update itself from a third party, and nothing on the device listens for inbound connections.

Delivery is off until the agent is pointed at your organisation. Before that, it can still scan, and nothing leaves the device.

## Light on the device

A scan is quick and uses little memory, and each report is a few tens of kilobytes. Scans are spread out slightly so that a fleet installed in one push does not report all at once. If the device is offline, reports wait on disk and are delivered when it reconnects.

Each report replaces the device's previous state rather than piling up. A tool that was uninstalled disappears from the console instead of lingering as a finding nobody can close. History is kept as changes over time, so you can still see when it was there.

## From scan to console

1. **The agent scans:** Every 15 minutes it reads installed AI tools, their settings, MCP servers, usage records, AI website visits and device posture. Secret values are discarded on the device.
2. **The report is delivered:** Over HTTPS to your organisation's EndLayer service. The device proves who it is with a credential of its own.
3. **The backend grades it:** Each signal is checked against your published policy, and the result keeps its evidence and the policy version behind it.
4. **The console shows it:** Findings appear in the console within moments, and anything that needs a decision goes to the Inbox.

## Connections fill the gaps

The agent sees the device. Connections see what a device cannot: managed devices that never enrolled, the people behind each device, and AI that runs in your cloud accounts and Microsoft tenant. Each connection reads your own account with a credential you provide. See [integrations](https://www.endlayer.ai/docs/connections).

  - [What is collected](https://www.endlayer.ai/docs/collection): Each signal group, and what is never read.
  - [Trust model](https://www.endlayer.ai/docs/trust): The promises built into the agent.

Source: https://www.endlayer.ai/docs/how-it-works · Updated 2026-11-20
