Agentplane is the Git-native control plane for coding agents

Use any repository-capable coding agent as the worker. Agentplane bounds its authority, holds approval gates, independently observes repository and Git facts, and records verification, recovery, and closure.

Compare layers, not logos

Layer	Primary job	What remains missing without Agentplane
Coding agents	Reason, edit, and test	The worker still reports on its own work
AGENTS.md and prompt policy	Describe rules	Instructions alone do not own lifecycle state or proof
Git and pull requests	Record content history and review	Git does not encode delegated authority or semantic intent
CI	Execute automated checks	A passing check does not explain scope, approval, or recovery
Fleet and operations platforms	Run and observe many agents	Operational breadth may not make the repository control boundary explicit
Agentplane	Control delegated coding work	Keeps authority, observation, and closure coherent across the other layers

Coding agents

Coding agents are responsible for reasoning and changing files. Agentplane is responsible for controlling the delegated lifecycle around them.

Use Agentplane when you need a reviewer to answer:

• What was the task?
• Which plan constrained the work?
• What did the supervisor independently observe?
• Which commit represents the work?
• Can the work safely close, or does it need recovery?

AGENTS.md and CLAUDE.md

Instruction files describe policy. They do not record whether a specific agent task followed that policy.

Agentplane keeps the policy gateway, then adds scoped authority, task state, plan approval, supervisor observation, verification, recovery, finish metadata, PR artifacts, and ACR.

Git-only workflows

Git remains the source of truth for content and history. Agentplane does not replace it.

The gap is that a commit hash does not describe delegated authority, semantic intent, approvals, verification, or recovery decisions. Agentplane stores those records beside the commit so review and revert have enough context.

CI

CI proves that configured checks ran. It usually does not know why the agent changed the files or which task evidence belongs to the pull request.

Agentplane can record verification as task evidence and use agentplane acr check <task-id> --mode ci as an ACR readiness gate.

Hosted agent platforms

Hosted runtimes and dashboards are useful when you need remote execution, central queues, or sandbox management.

Agentplane OSS CLI is not a hosted agent runtime or sandbox API. It is a local-first control plane for coding-agent work in repositories.

Fleet-first platforms are strongest when the first job is standing up many workers, connecting adapters, and operating them through a dedicated UI. Agentplane is strongest when the first job is making delegated authority and proof explicit inside the repository across any coding agent.

When to use Agentplane

Use Agentplane when:

• your coding agent changes real repositories;
• reviewers need authority, intent, verification, recovery, and closure evidence;
• AI-written code changes need a repeatable review contract;
• you want durable artifacts in Git instead of transient chat history.

Do not use Agentplane when:

• you only need a prompt playground;
• you want a hosted sandbox runtime;
• you want an agent framework to choose models and tools;
• you do not use Git-based review.

Agent-agnostic by design

Agentplane does not depend on a specific model, editor, hosted runtime, or agent product. Git stores the diff. CI runs checks. Agentplane records the evidence around the work.

Control-plane distinction

The worker's semantic result and the control plane's observed facts are separate records. Agent confidence does not silently become supervisor proof.

Let agents write code. Keep authority and proof in Git.