Skip to main content

Agents

The agent system is the execution-heavy half of II-Agent. It handles multi-step tasks, tool use, file mutation, session continuity, and streamed feedback to the frontend.

Major Runtime Areas​

  • agents/models/ for provider/model abstractions
  • agents/runs/ for run persistence
  • agents/sandboxes/ for isolated execution environments
  • agents/skills/ for built-in and custom skills
  • agents/tools/ for executable tool families
  • agents/plans/ and prompts for structured agent behavior

Tool Families​

The current tool tree includes:

  • agent
  • browser
  • connectors
  • dev
  • file_system
  • media
  • plan
  • productivity
  • sandbox
  • shell
  • slide_system
  • web

That breadth is why the agent docs need more than “it can call tools.”

Session and Run Model​

Agent work sits on top of:

  • sessions
  • run tasks
  • agent run tasks
  • agent run messages
  • application events

This gives the UI enough data to replay, inspect, and continue work instead of treating execution as a black box.

Skills​

The platform supports:

  • built-in skills checked into the repo
  • user-configurable/custom skill settings
  • specialized execution helpers such as browser or document skills

Skills are an important part of the extension story because they let runtime behavior evolve without every capability becoming core application code.

Sandboxes​

The repo is set up to work with sandbox providers and related server components. In practice that means:

  • execution timeouts matter
  • provider keys and template IDs matter
  • docs must explain both the runtime and the env model clearly

When To Read This Page​

Use this page when you need to understand:

  • where agent behavior lives
  • which tool family should own a new capability
  • how streamed execution becomes persisted state
  • how agent mode differs from chat mode in practice