I've just released vincent 0.12.0, the biggest update so far to my open-source AI SDLC framework.

For those unfamiliar with the project, vincent is a local-first, vendor-independent control plane for AI coding agents. It supports Claude Code, Codex, Cursor, and now GitHub Copilot CLI.

The idea behind it is fairly simple: AI agents are getting increasingly capable of writing code, but generating code is only one part of software engineering.

We still need to manage tasks, coordinate changes, run tests, review results, handle failures, and decide what actually gets merged.

That's the problem I'm trying to solve with vincent.

What's new in 0.12.0?

The biggest change is a refactor of the development process into three distinct stages:

Intake: Issues are managed directly within vincent, including synchronization with GitHub. Work no longer needs to begin with manually constructing prompts.

Workshop: Agents work on issues using isolated Git worktrees, structured workflows, and verification steps. Multiple tasks can contribute to the same issue without stepping on each other's changes.

Handoff: Completed work is delivered through separate integration workflows, with explicit human authorization for operations such as creating or merging pull requests.

The separation matters. An agent finishing its task shouldn't automatically mean its changes are ready for production.

This release also adds GitHub Copilot CLI as the fourth supported coding agent, an improved issue board, better task orchestration, and several usability improvements.

Using vincent to develop vincent

One of the more interesting aspects of this project is that I increasingly use vincent to develop itself.

Between versions 0.11.0 and 0.12.0:

  • 765 commits
  • 73 merged pull requests
  • 64 pull requests created from vincent-managed task branches

It's a useful way to test the framework against actual development work rather than relying entirely on artificial examples.

It also exposes problems pretty quickly. When you're coordinating multiple agents, handling concurrent changes and reviewing their output, you start noticing that the bottleneck isn't necessarily writing code anymore.

It's managing everything around it.

Why I'm building this

I've been developing software for over two decades, and I see enormous potential in AI-assisted development.

But I don't believe the solution is simply giving agents more autonomy and hoping for the best.

I'd rather have a system where agents can do substantial amounts of work while the development process remains understandable, reproducible, and under human control.

That's what I'm working toward with vincent.

Version 0.12.0 is a significant step in that direction, although there's still plenty left to improve.

I've published a detailed release article covering the new architecture, workflows, CLI commands, and a complete example going from a GitHub issue to a merged pull request.

You can read it here:

https://blog.lezli01.is-a.dev/blog/vincent-0-12-release/

The project is open source, and I'd be interested to hear how others are approaching orchestration, verification, and human oversight in agentic development workflows.