Flagship project
Kommandozentrale — AI Agent Orchestration Platform
My own platform for running several AI coding agents on one codebase: every task in its own branch, every change checked before it reaches the main line, every step on record. It is in daily use in my own development work. The goal is a product for development teams.
The problem
Several AI agents working on one codebase go wrong in quiet ways. Two agents overwrite each other's work. Nobody knows which state is the current one. A change that looks fine in one session breaks something in another.
More agents do not make that easier to control. They only make the lack of control show up faster.
The solution
The Kommandozentrale turns agent work into a controlled process. Each task gets its own branch and its own working copy. Agents never write to the main line directly. A merge gate checks every change before it is accepted, and every merge leaves a record of what was checked. Deploying and publishing stay with me.
- Task — the work is defined and gets its own branch and an isolated working copy.
- AI agent — Claude Code (Codex and local models are connected too) works inside that copy and nowhere else.
- Tools & execution — scripts, skills and safeguards run around the agent. Commands that could destroy something or leak a secret are stopped first.
- Testing — the change is tested in a throw-away copy of the project, not in the live one.
- Review gate — the merge gate looks for secrets, accidental deletions and failing tests. What fails goes back; nothing is merged half-checked.
- Result — the change reaches the main line together with a record of what was checked.
Starting several agents is easy. Keeping their work under control is the hard part.
What it can do today
Status as of 10 October 2026. “In use” means used in my own daily work. “Tested” means shown to work in tests but not run continuously.
- In use
Tasks in isolated branches
Every task runs in its own branch and its own working copy. The main line is closed to agents.
- In use
Merge gate
A fixed sequence of checks before anything is merged: a secret scan, a trial merge, a guard against accidental loss and mass deletion, and the tests in a throw-away copy. A conflict produces a report instead of a blind merge.
- In use
Several agents side by side
Several Claude Code sessions work on the same codebase at the same time, each in its own copy, also across several computers. Codex is connected as a second agent.
- Tested
Several computers
Agents on different computers hand tasks to each other through a local message bus. Tested end to end between machines.
- Tested
Local AI models
Local models, run through LM Studio, are connected and have been measured against cloud models in comparisons. Not run continuously.
- In use
Safeguards before risky commands
Hooks stop deployments, destructive deletions and secrets in commits before they run. Built on a guard package I integrated, plus hooks of my own.
- In use
Traceability
Every merge carries a record of the checks it passed. A journal and a task log show who did what, and when.
- In use
Reusable tools, skills and workflows
Recurring work lives as scripts, skills and workflows instead of being explained again each time.
- In development
Control center
A local dashboard with a system map and a project overview. It runs on my own machine only.
How the parts fit together
The agents are interchangeable workers. Around them sit the parts that carry the control: task branches and working copies keep the work apart, tools and hooks limit what an agent may do, test runs show whether a change holds, and the merge gate decides what reaches the main line. People release: an agent never deploys or publishes on its own.
- Task branches
- Working copies
- Agent providers
- Hooks & safeguards
- Test runs
- Merge gate
- Audit record
In operation
- 190
- merges accepted by the gate
- 208
- changes submitted to it
- 14
- failed a check at least once and were sent back
Counted from the gate's own log, 29 September to 10 October 2026, 20:00. Every figure was counted twice with differently built methods, and the 190 was also checked against the repository history. The log keeps growing, so these numbers are a snapshot.
Technical challenges and how they are solved
- Agents overwriting each other: every task gets its own branch and working copy, and the main line accepts nothing directly.
- Quality before integration: every change goes through automatic checks, so “looks fine” is not the standard.
- Knowing what happened: each merge records its checks, so a day of agent work can be reconstructed afterwards.
- Several model providers: agents are interchangeable workers behind the same process, so the control does not depend on one vendor.
Technologies
- Git
- Node.js
- Python
- Claude Code
- Codex
- Local models
Current status
The Kommandozentrale is an internal development system. I use it every day in my own work. It is not yet a product for third parties: there are no outside users, no pricing and no release date.
Where it is going
The aim is to turn the internal system into a product for development teams that run several AI coding agents on one codebase: software teams, small studios and individual developers who work with more than one agent.
This is a direction, not a result. There are no customers yet, and nothing here is a promise.
Perspective: Claude
Claude Code is one of the agents the system drives today. A deeper Claude integration could bring more developers and companies to use Claude models through the Kommandozentrale: a team that coordinates its agents here would run its Claude work through the same controlled process.
That is a prospect, not a result. There are no usage figures to show, and this page makes no claim about any partnership.