# AgentPlane: Git-Native Governance Framework for Deterministic and Auditable AI Agent Workflows

> AgentPlane is a CLI tool that runs locally on Git repositories, providing explicit governance mechanisms for AI agent-driven development workflows. It enables auditable and trustworthy agent collaboration through visible artifacts such as task statuses, verification records, and deterministic closed loops.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-04-03T16:16:46.000Z
- 最近活动: 2026-04-03T16:23:05.265Z
- 热度: 163.9
- 关键词: AgentPlane, AI代理, Git工作流, 代码治理, CLI工具, 确定性工作流, 可审计, 代理协作, 软件开发, 任务管理
- 页面链接: https://www.zingnex.cn/en/forum/thread/agentplane-aigit
- Canonical: https://www.zingnex.cn/forum/thread/agentplane-aigit
- Markdown 来源: floors_fallback

---

## Introduction / Main Floor: AgentPlane: Git-Native Governance Framework for Deterministic and Auditable AI Agent Workflows

AgentPlane is a CLI tool that runs locally on Git repositories, providing explicit governance mechanisms for AI agent-driven development workflows. It enables auditable and trustworthy agent collaboration through visible artifacts such as task statuses, verification records, and deterministic closed loops.

## Background: Trust Dilemma in AI Agent Development

With the popularity of AI coding assistants and code agents, development teams face a core challenge: how to maintain control and auditability over code changes while enjoying the efficiency gains from automation. Traditional AI assistant models are often "black-box"—users make requests, agents modify files directly, and the entire process lacks structured records and clear approval nodes. This model is acceptable for individual rapid iteration scenarios, but in environments with team collaboration, strict code reviews, and high compliance requirements, the practice of "letting agents modify files directly" makes it difficult to build trust. Teams need to know: What did the agent do? Why did it do that? Who approved these changes? How to verify the correctness of the results?

## AgentPlane's Solution Approach

AgentPlane proposes a completely different paradigm—integrating AI agent workflows into the Git governance system. It does not aim to replace Git, editors, or terminals; instead, it embeds a layer of visible agent workflow operation models into existing repositories, making agent execution inspectable, auditable, and rollbackable. The core idea can be summarized as: **Trust comes from visible states**. All approvals, task state transitions, verifications, and completion records are stored in the repository, not in the black box of an external hosting platform.

## Core Architecture and Artifact System

AgentPlane introduces a set of explicit artifacts into the repository, making agent work readable for both humans and agents:

## Policy Gateway Files (AGENTS.md / CLAUDE.md)

Policy files located in the repository root directory define the agent work policies for the repository. This includes agent role definitions, work modes, approval processes, code style requirements, etc. It serves as both an entry point for human reviewers to understand agent work rules and a source of behavioral guidelines for the agents themselves.

## Agent Workspace (.agentplane/)

The local agent workspace directory in the repository includes:
- **config.json**: Current repository configuration state
- **tasks/**: Task-organized record directory
- **WORKFLOW.md**: Materialized workflow contract describing current workflow modes and rules
- **tasks.json**: Optional task snapshot export

The key advantages of this design are **locality and inspectability**. There is no external hosted control plane involved; all states are stored in the Git system familiar to developers.

## Workflow Modes

AgentPlane supports two integration styles to adapt to different collaboration needs:

**Direct Mode (Default)**:
- Single checkout repository
- Rapid iteration in the current working tree
- Suitable for individual development and quick prototyping
- Creates deterministic closed-loop commits by default when tasks are completed

**Branch_PR Mode**:
- Structured task-based branch or working tree process
- Generates explicit PR artifacts under `.agentplane/tasks/<task-id>/pr/`
- Suitable for teams requiring strict handover and integration discipline
- Seamlessly integrates with existing code review processes

## Task Lifecycle and Governance Mechanism

AgentPlane models agent work as an explicit task lifecycle, with clear records and approval requirements for each stage: