---
title: "Key Concepts"
description: "Before diving into Nexus OS, understand these fundamental concepts that appear throughout the documentation. An agent is a self contained unit of work. It receives input, performs a task, and produces output. Agents run "
resource: https://www.aiagents.nexus/docs/manual/introduction/key-concepts
generated: { by: "process:nexus-agent-assets", at: 2026-09-07T09:13:03Z }
status: stable
---

# Key Concepts

Before diving into Nexus OS, understand these fundamental concepts that appear throughout the documentation.

## Agent

An agent is a self-contained unit of work. It receives input, performs a task, and produces output. Agents run in WASM sandboxes for isolation and security.

```bash
naos create my-agent --template researcher
```

## Orchestration Primitives

Nexus provides four ways to coordinate agents:

| Primitive | Purpose | Analogy |
|-----------|---------|---------|
| Supervisor | Restart failed agents | A manager who rehires when someone quits |
| Saga | Multi-step with rollback | A transaction that undoes itself on failure |
| Workflow | Sequential pipeline | An assembly line where each station passes to the next |
| Pool | Parallel execution | A team working on the same task simultaneously |

## Supervision Tree

Supervisors can supervise other supervisors, creating a tree:

```
root-supervisor
├── api-supervisor
│   ├── auth-agent
│   └── data-agent
└── worker-supervisor
    ├── processor-1
    ├── processor-2
    └── processor-3
```

If `processor-1` crashes, `worker-supervisor` restarts it. If `worker-supervisor` itself crashes, `root-supervisor` restarts it and all its children.

## Cost Budget

Every agent can have a spending limit:

```yaml
cost:
  budgets:
    researcher:
      limit: 1000    # cents ($10.00)
      period: day
      onExceed: pause
```

## Trust Score

AXIS Trust assigns scores to agents:
- **T-Score**: 0-100 (overall trustworthiness)
- **Trust Tier**: T1-T5 (classification)
- **Credit Rating**: AAA to D (transaction reliability)

## WASM Sandbox

Agents compile to WebAssembly and run in an isolated sandbox. No file system access, no network access, no ambient authority — only what you explicitly grant.

## Audit Log

Every action in Nexus OS is logged:

```bash
naos audit --last 10

# [2025-01-15 10:30:01] agent.created    researcher
# [2025-01-15 10:30:02] agent.started    researcher
# [2025-01-15 10:31:15] cost.recorded    researcher  $0.05
# [2025-01-15 10:31:16] agent.completed  researcher
```

## Project

A Nexus project is a directory containing:
- `nexus.config.yaml` — Configuration
- `data/` — SQLite database
- `agents/` — Agent source files
- `skills/` — Skill definitions