# How AgentGit compares

Honest positioning, updated as the landscape changes. If something here is
outdated or wrong, please open an issue.

| | Scope | Isolation | Validation before merge | Runtime lock-in | Maturity (as of 2026-07) |
|---|---|---|---|---|---|
| **AgentGit** | Task decomposition entry point → branch isolation → deterministic validation → automated merge | Branch/worktree per task, enforced | Core feature — a real command must exit 0 | None (CLI is Node, but the file format needs only text + git) | Pre-alpha, MVP |
| **gnap** | Coordination protocol only (who owns what task) | Not specified by the protocol | Explicitly out of scope — left to the application layer | None (plain files) | Early, in limited production use |
| **CrewAI / LangGraph** | Agent orchestration as a Python library | In-process, not git-based | Application-defined | Python | Mature, widely used |
| **Worktree session managers** (e.g. Claude Squad, Conductor-style tools) | Run/manage multiple coding-agent sessions in parallel worktrees | Worktree per session | Usually human-reviewed, not an automated gate | Varies by tool | Mature, actively used |

## The honest gap

Everything in that last column matters: most of the tools in this table are
further along than AgentGit is today. What AgentGit is betting on is a
specific, narrow claim — that the validation-before-merge gate should be a
deterministic script, not a human glance or an agent's self-report — and
building that one piece well before expanding scope. If your need is solved
by a more mature tool in this table today, use that tool.