​
Overview

Native Agent is the primary chat surface for running autonomous agents inside the Allternit platform. It combines a conversational message stream with a live execution canvas so you can watch an agent think, plan, call tools, and report progress in real time. Unlike the general chat view, Native Agent sessions are isolated by origin surface (chat), load agent workspace context, and surface structured agent output such as milestone roadmaps and tool execution traces.

​
How to access it

Native Agent runs inside the platform shell. Open it from the shell navigation or programmatically:
From shell code:
The surface is registered as a singleton view (native-agent), so only one instance can be open at a time.

​
Features

Session Management

Create, rename, switch, and delete agent sessions from a single dropdown

Split Chat + Canvas

Toggle between split, chat-only, and canvas-only layouts

Streaming Messages

Real-time assistant responses with thinking indicators and tool-call events

Milestone Progress

Visual roadmap of agent missions, waves, and completed tasks

Tool Call Visualization

Expandable cards for every tool call, argument, result, and error

Runtime Ops

One-click access to runtime controls, execution mode, and backend status

​
Layout modes

Use the control bar in the header to change how the chat and canvas are arranged.

​
Key concepts

​
Session

A Native Agent session is a persistent conversation tied to the chat origin surface. Sessions are stored through the backend at /api/v1/agent-sessions and can operate in either regular or agent mode. Agent-mode sessions load an agent workspace and attach its context pack to every message.

​
Context pack

When a session runs in agent mode, the platform builds a context pack from the agent’s workspace files: The pack is cached by content hash and refreshed automatically when workspace files change.

​
Canvas

The right-hand canvas visualizes the agent’s execution state. It renders two primary components:
  • MilestoneProgress — a linear roadmap of milestones plus the active mission’s task path.
  • ToolCallVisualization — a list of tool calls with arguments, results, execution times, and retry actions.

​
Tool calls

Tool calls are rendered as collapsible cards grouped by category (File System, System, Code, Automation, Interactive). Each card shows:
  • Tool name and category
  • Live execution status
  • Arguments
  • Result or error output
  • Execution time

​
Runtime Ops

The Runtime Ops button in the header opens the runtime operations view, where you can inspect execution mode, backend health, and active sessions.

​
Endpoints

Native Agent uses the N20 Native Agent API client (native-agent-api). Key routes:

​
Session modes

Set the mode when creating a session via session_mode in the request payload, or programmatically through the mode session store.

​
Runtime execution modes

The runtime API supports three execution modes that control how aggressively the agent acts: Change the mode through Runtime Ops or by calling PUT /api/v1/runtime/execution-mode.