It's in active development in my free time. I've built various agent orchestration systems over the years for different reasons, ever since the GPT-3 API. I can tell you utility has continually risen, the models are just getting better, and late 2025 was an inflection point, which is why we're seeing all of these orchestration solutions pop up now.
I still have kinks to work out in mine but it's already usable for building software. Once I get to v1 I think it will provide enough value to be useful for me in particular. I don't have enough data to speak about months on yet, but if I think the experiment is a success then I will do a Show HN or something.
The gist is you can clone a repo or start a project from scratch, each engineering agent gets a worktree, you work with the manager agent and it dispatches and manages other agents. there are playbooks which agents contextually turn into specific tasks, each of which is tracked much like CI/CD. You can see all the tool calls, and all of the communication between both agents and humans.
The application model is ticket-based. Everything revolves around the all-holy ticket. It's like a prompt, but it becomes a foundation for tying together every bit of work related to the process of developing the feature. So you can see the progress of the ticket through the organization kanban style, or watch from a dashboard, or look at specific tickets.
There are multiple review steps where human review and intervention are required. Agents are able to escalate to humans whenever they think they need to. There is a permission system, where agents have to seek permissions from other agents or humans in a chain of command in order to do certain tasks. Everything is audited and memoized, allowing for extreme accountability and process refinement stages.
Additionally, every agent "belongs" to either another agent or a human, so there is always a human somewhere in the chain of command who is responsible and accountable for the actions of his agent team. This team includes the manager agent, engineering agents, test agents, QA agents, etc, each loaded with different context, motivations and tools to keep them on track and attempt to minimize the common failure modes I experience while working closely with these tools all day.
I still have kinks to work out in mine but it's already usable for building software. Once I get to v1 I think it will provide enough value to be useful for me in particular. I don't have enough data to speak about months on yet, but if I think the experiment is a success then I will do a Show HN or something.
The gist is you can clone a repo or start a project from scratch, each engineering agent gets a worktree, you work with the manager agent and it dispatches and manages other agents. there are playbooks which agents contextually turn into specific tasks, each of which is tracked much like CI/CD. You can see all the tool calls, and all of the communication between both agents and humans.
The application model is ticket-based. Everything revolves around the all-holy ticket. It's like a prompt, but it becomes a foundation for tying together every bit of work related to the process of developing the feature. So you can see the progress of the ticket through the organization kanban style, or watch from a dashboard, or look at specific tickets.
There are multiple review steps where human review and intervention are required. Agents are able to escalate to humans whenever they think they need to. There is a permission system, where agents have to seek permissions from other agents or humans in a chain of command in order to do certain tasks. Everything is audited and memoized, allowing for extreme accountability and process refinement stages.
Additionally, every agent "belongs" to either another agent or a human, so there is always a human somewhere in the chain of command who is responsible and accountable for the actions of his agent team. This team includes the manager agent, engineering agents, test agents, QA agents, etc, each loaded with different context, motivations and tools to keep them on track and attempt to minimize the common failure modes I experience while working closely with these tools all day.