玄衡科技XUANHENG TECHNOLOGY
MENU

PROJECT 02 · AGENT SYSTEM

Agent Workspace

A local, single-user native agent workspace where coding agents run under a local-terminal execution contract.

Status: Research Prototype

FOCUS

  • AI Agent
  • Human-AI Collaboration
  • Task Continuity
A live native session in a local demo workspace: the agent reads the repository, reports uncommitted changes and summarizes its structure inside the terminal.
A live native session (Claude Code) captured from a local demo workspace; the in-app brand reads Xuan.

01 · Overview

Overview

Agent Workspace is a local, single-user native agent workspace. Codex, Claude Code and Antigravity appear in its main area exactly as their own native interfaces; StepFun is the workspace’s own provider, calling its model API directly through a built-in session. The app takes care of workspaces, sessions and process management.

CLI sessions follow the local-terminal contract: they inherit your environment and configuration without extra wrapping or restrictions, and StepFun tool calls run under the same permission and evidence rules. A session survives view switches and refreshes, and the app ends it only when you confirm; unexpected exits are recorded as they are observed.

For long-running work, an optional task handoff records the goal, authorizations, verified state and acceptance criteria as an immutable document; after review, a fresh session continues under the same contract — an advanced feature, never a burden on daily use.

02 · Capabilities

Capabilities

  1. 01

    Native Agent Sessions

    Codex, Claude Code and Antigravity in their own TUIs, plus a first-party StepFun session on a direct API.

  2. 02

    Local Terminal Contract

    CLI sessions inherit your environment and configuration; when a capability is missing, launch is refused rather than quietly degraded.

  3. 03

    A Full Session Lifecycle

    Create, rename, switch, reconnect, take over across tabs and end explicitly, with process state always visible.

  4. 04

    Task Handoff & Continuation

    Goals, authorizations and acceptance criteria recorded as immutable documents, reviewed before a new session picks them up.

  5. 05

    External Integrations

    Controlled task execution and read-only project access for automation and tooling, under the same permission boundaries.

03 · How It Works

How It Works

  1. 01

    Pick a workspace

  2. 02

    Choose a provider and open a session

  3. 03

    Work the way you would in a terminal

  4. 04

    Switch, reconnect or end

  5. 05

    Optional: hand off a task to continue

04 · System View

System View

  1. 01

    Interaction

  2. 02

    Provider

  3. 03

    Execution

  4. 04

    Governance

How it works: five daily steps from picking a workspace to ending a session, with task handoff drawn as an optional branch, above the four interaction, provider, execution and governance layers.
The daily path, the optional handoff branch and the four layers; as of September 2026.

05 · Current State

Current State

Implemented

All four providers are available today: native sessions, the full session lifecycle with tab takeover, read-only historical chats, task handoff and continuation, external integrations, and resource coordination for concurrent work.

Current Focus

Making long-running sessions more dependable — reconnection, crash recovery and resource coordination.

Next Direction

Keep native sessions at the center while refining task continuity and status observability.

06 · Principles

Principles

  1. 01

    Native sessions are the product; the app presents and manages them.

  2. 02

    Consistent with the local terminal: when something cannot run, it says so instead of quietly degrading.

  3. 03

    Session and task state is based on real process facts, never guesses at screen text.

  4. 04

    Handoff and integrations follow the same permissions and records — no second system.