The graph that learns while you ship

A static graph is a map. A self-learning graph is a memory.

Helix runs five deterministic agent loops on the Engineering Graph. They do not invent code. They do not guess. They observe evidenced relationships and write patterns back onto the graph so the next question starts from something the team already proved.

The loops

• Silo finds files with a single author and encodes the bus-factor risk.
• Coupling notices files that repeatedly land in the same pull requests.
• Expertise weights who actually knows a path, recency first.
• DNA preserves how this team reviews the hottest code.
• Memory reinforces the nodes future answers should start from.

Each cycle is one loop. Each loop produces a pattern node, WATCHES edges from the agent, and LEARNED edges to the people and files involved. You can pause it. You can click the pattern. You can ask Helix about it.

const result = applyCycle(graph, generation);
graph = result.graph;
// generation 12 stays 12 if you pause
// resume starts at 13

The generation count is part of the graph, not a toy in the browser. Reload the workspace and Helix continues from the same generation. Pause, and it waits. It does not reset because you closed a laptop.

Deterministic before generative

The loops prefer deterministic reasoning before AI reasoning. Facts stay facts: authors, reviewers, files, timestamps. Inference is labeled as a pattern, with a generation number, so you can see when the graph learned it.

That is how software starts to explain itself. Not by searching a wiki. By compounding every merge into Engineering DNA.

A loop that cannot be paused is automation without consent. A loop that forgets its generation is not learning. Helix does both: you stay in control, and the memory stays.

Watch a sample graph encode DNA in a few generations, then index your own. Open Helix: learning stays on until you pause it.