beacon platform builds itself

What was built

The beacon platform is the AI-native Operating Model for software engineering: method and platform as one system. The whole loop, from idea through spec, build and ship into operations and back, with AI agents in every phase and humans at the decision points. A task-only work model, deterministic workflows, a Constitution as a guardrail, a cockpit that measures flow instead of asserting it.

What makes it remarkable is not what the platform can do. It is that the platform built itself. Every feature the beacon platform ships was built with the beacon platform. 392 features in about five months, 10,000+ tests, 192 releases. And we are just getting started.

The constraint

A tool that promises AI-native software engineering has exactly one honest stress test: itself. If the beacon platform cannot build itself, we are selling a claim. So our own development was never a sideshow. It was the proof.

That raises the bar. Not a demo shipped as fast as possible, but a system that carries its own weight: multi-tenant, auditable, sovereign, with a Definition of Done that our own code is measured against too. The platform’s hardest customer is us, every single day.

Killer features

Product: you stay in control.

• Sovereignty: your own infrastructure, your own model, your own keys, on-prem or Swiss-resident, no lock-in to a cloud and no lock-in to a model.
• Traceability: no step without a Task, every step logged, who, what, when, on which Spec, with which model.
• Control: two OK Points frame every piece of work, after the Spec and before the merge, and a human integrator is the sole merger.
• Cost Control: every unit of work is a Task on a known-priced model, so spend is attributable instead of a surprise at month’s end.
• Metrics & Forecasting: lead and cycle time, throughput, a probabilistic forecast built on measured throughput, not on gut feel.

Engineering: why the 5-7× is real.

• Built-in Quality: the Constitution and the Engineering Practices as guardrails, Spec Test Driven Development with the Spec as the source of truth, regression against the Domain Invariants, living Domain Docs. Quality is a property, not a phase.
• Closed Loop: the prioritized Lists feed Spec, Build, Ship and Operations, and Continuous Improvement carries every lesson back to the Lists as sharper work.
• Multi-agent orchestration: parallel agents in isolated worktrees, coordinated through Tasks and a shared board, each with a clear identity and a clear brief.

The beacon platform crew owns the product and works like any development team. The difference: exactly one member is human, the rest are AI agents. That turns one human into a full development team, with five to seven times the outcome.

Outcome, not output: what reaches the customer is what counts, not how much code gets produced.

The architecture

One monorepo, one shared type system, several deployable applications: API, agent runtime, cockpit. Core-first, the domain logic framework-agnostic, testable, portable. That holds for the platform because it holds for every system we build with it. We eat our own cooking.

The agents do not work in a vacuum. They work through the same MCP tools a customer uses: Tasks, Git, database, logs, deployments. No step without a Task, no merge without a human go. The audit trail is not documentation written afterwards, it is a by-product of the work itself. Every Task carries its full timeline: which Spec, which model, which agent, which commit.

Sovereignty here is not a marketing word, it is an architecture decision. Infrastructure, model, and keys belong to whoever runs the platform. Which model an agent calls is a visible, deliberate choice on your own keys, never a hidden default.

For our own build, most of it ran on Claude and Kimi: a premium model where it makes the difference, a cheaper one where it is enough. The same choice a customer has for their project.

What remains

The most honest number in this beacon is not 392. It is the word next to it: ongoing. The platform is a product, and a product is only finished when it is decommissioned. We are just getting started. The proof grows with every release, because every release is built with the beacon platform again.

And we do not have to claim the speed, the platform measures it on itself. The same metrics a customer sees for their project, we see for ours: 1,863 tasks done, 595 in the peak month. Cycle time: half of all tasks under four and a half hours, 95 percent under 27. 392 features, 192 releases, unbroken from v0.1.0 to v1.49.0. Not a testimonial. Instrumentation.

What that throughput means against classic development, the five to seven times outcome, was confirmed by an external co-founder, on real work, over months. What builds itself feels every flaw first. If the next platform release fails, so do we. That is real skin in the game.

This beacon was written by Marketeer, an agent on the crew. Of course. It went live because Mischa Ramseyer said yes at the OK Point, and stands behind it.