Skip to main content

Why coding agents need a Git-native control plane

Coding agents made code generation cheap. They did not make delegated change accountable.

Agentplane is the Git-native control plane for coding agents. It exists so teams can let agents write code while keeping authority and proof in Git.

1. The failure mode​

AI agents can now edit real repositories, not just draft snippets. They update code, docs, tests, release paths, and configuration. That creates a new review problem: the repository shows the final diff, but the reasoning trail often remains in a chat transcript, terminal scrollback, or IDE state.

The weak link is not only "can the agent write code?" It is whether the team can prove what was authorized, what the supervisor observed, and why closure was justified.

2. The worker does not grade itself​

The coding agent is a semantic worker. It can reason, edit, test, and report a result. Its report is valuable, but it remains a claim by the worker.

Agentplane owns the lifecycle around that worker. It issues bounded work, holds approval boundaries, observes repository and Git state, records verification, and determines the next safe route. Agent-reported facts do not silently become supervisor-observed facts.

3. Why Git alone is not enough​

Git is the right durable substrate. It stores content, branches, commits, and review history. But a commit does not reliably answer:

  • What was the agent asked to do?
  • Which plan constrained the change?
  • Which checks actually ran?
  • Which policy decisions or waivers applied?
  • Why is the work safe to merge or revert?

Agentplane keeps Git as the source of truth and adds the missing evidence beside the diff.

4. Why chat transcripts are not enough​

Chat transcripts are useful context, but they are poor review artifacts. They are long, mutable, private, tool-specific, and often unavailable to the next reviewer. They can include sensitive material that should not be committed.

Agentplane records commit-safe summaries, task state, verification evidence, and hashes instead of requiring reviewers to trust raw agent conversations.

5. Why the durable record belongs with the code​

Remote execution and central status can be useful. They are separate from the durable evidence record that reviewers need next to the code.

Agentplane keeps the durable workflow record with the repository. The repository owns the workflow artifacts.

6. Why ACR exists​

An Agent Change Record is the evidence file for AI-authored work. It is a commit-safe JSON artifact that records task intent, approved plan, agent/toolchain identity, Git commits, policy decisions, verification results, and merge readiness.

Humans can read agentplane acr explain. CI can run agentplane acr check. External tools can read .agentplane/tasks/<task-id>/acr.json.

7. What Agentplane refuses to do​

Agentplane is not another coding agent. It does not choose a model, generate code for you, replace Git, replace CI, or replace pull request review.

It wraps the tools you already use with a repo-local lifecycle:

authorize -> dispatch -> observe -> verify -> close or recover

8. The operating bet​

The durable interface for agentic engineering will be a protocol for delegated authority and observed proof that any capable coding agent can follow. That is the category Agentplane intends to own.