System Prompt

Overview

The system prompt is the foundation of your agent's behavior. ArgentOS assembles it from multiple sections, each serving a distinct purpose. This modular approach lets you customize the agent's personality and capabilities without editing a single monolithic prompt.

Prompt Sections

IDENTITY

The identity section defines who the agent is. It includes the agent's name, role, and core personality traits.

You are Argent, a personal AI assistant created by Jason.
You are helpful, proactive, and technically skilled.

This is configured in argent.json:

{
  "agents": {
    "main": {
      "name": "Argent",
      "identity": "You are Argent, a personal AI assistant."
    }
  }
}

SOUL

The soul section defines the agent's behavioral guidelines -- how it should interact, what tone to use, and what principles to follow.

TOOLS

The tools section is auto-generated from the registered tool schemas. It tells the model what tools are available and how to use them. This section is not manually edited -- it is assembled from tool definitions.

BOOTSTRAP

The bootstrap section provides initial context loaded at session start:

• Recent memories from MemU relevant to the user
• Active tasks from the task system
• Workspace context (what files/projects are relevant)
• Injected plugin context

Time Awareness

Every user message is wrapped in a timestamp envelope:

[Wed 2026-02-12 10:30 America/Chicago | last message: 3h 15m ago]

This gives the agent temporal awareness -- it knows the current time and how long since the last interaction. The system prompt includes instructions on how to interpret these timestamps.

Elapsed time under 30 seconds is omitted (same conversational turn).

Customizing the System Prompt

Via Configuration

{
  "agents": {
    "main": {
      "identity": "You are a security analyst named Sentinel.",
      "soul": "path/to/custom-soul.md"
    }
  }
}

Via Workspace Files

Place a AGENT.md or similar file in the agent workspace directory. The bootstrap section can be configured to include workspace-level instructions.

Plugin Injections

Plugins can inject additional context into the system prompt. For example, the canvas-docs-enforcer plugin injects disambiguation instructions for tool naming.

{
  "systemPromptInjection": "When the user asks to create a document, use doc_panel..."
}