> ## Documentation Index
> Fetch the complete documentation index at: https://docs.amnify.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# How the agents work

> Specialists for each security topic, and a lead that connects what they find.

Amnify is not one model answering questions. It is a team, organized the way a good security team is: specialists who go deep in one area, and a lead who holds the whole picture.

```mermaid theme={null}
flowchart LR
  S1["Code and dependencies"] --> A(("Amy<br/>Security Lead"))
  S2["Infrastructure and cloud"] --> A
  S3["Live attack surface"] --> A
  S4["Your knowledge<br/>and decisions"] --> A
  A --> O["One picture:<br/>ranked findings,<br/>questions and fixes"]
```

## Specialists for each topic

Every security topic has its own agent, equipped with the skills, instructions and playbooks for that topic. That is what makes it a specialist rather than a generalist working from a checklist. It knows how an issue in its area is actually exploited, what normally mitigates it, and what it exposes when it is real.

Because it is a specialist, it is trusted to judge its own area: whether a result is a genuine weakness, whether an attacker could reach it, and what it would cost you if they did. Specialists work in parallel across code, dependencies, infrastructure as code, cloud and your live attack surface.

## Amy connects the dots

A specialist sees its own slice. **Amy**, the Security Lead, sees all of them at once, together with everything Amnify knows about how your organization runs.

Working with a small number of coordinating agents, Amy:

<CardGroup cols={2}>
  <Card title="Brings it into one picture" icon="layer-group">
    Separate reports from separate areas become a single, deduplicated view rather than five tool outputs to reconcile.
  </Card>

  <Card title="Finds the shared cause" icon="route">
    Recognizes when issues in different areas are the same underlying problem, so one fix closes all of them.
  </Card>

  <Card title="Decides what matters now" icon="gauge">
    Weighs each issue against your context and your decisions, and puts the queue in an order that reflects your organization.
  </Card>

  <Card title="Knows when to ask" icon="circle-question">
    When the call is yours rather than Amnify's, it asks instead of guessing. See [Context](/knowledge/context).
  </Card>
</CardGroup>

That end to end view is the part a single scanner cannot have. Deduplication, ranking and root cause analysis all need to see everything at once, not one area at a time.

## Nothing is taken on trust

Everything that reaches an agent is input to be judged, never an instruction to follow. That holds wherever it came from: a file in a connected repository, a document you uploaded, a decision already in your knowledge base, a comment on a pull request, or a page encountered during a test.

When any of it conflicts with security best practice, Amnify does not quietly comply. It flags the conflict, shows what the conflict is based on, and asks a person to confirm before it is applied. Your organization can still decide against best practice, and often will for good reasons, but that decision is made deliberately and recorded, never inherited in silence from a document or a source.

## Nothing is hidden

Every piece of work happens in a run that keeps its full transcript, so any conclusion can be traced back to the steps that produced it. See [Agent runs](/insight/agent-runs).