A genome you can audit

Emergence is a story you tell after the fact. A genome is a file you can diff. If we are going to let agents evolve, the lineage has to be as inspectable as the output.

The temptation in agent design is to hide the interesting parts. A large prompt, a bag of tools, a memory store, and a claim that the system 'learns'. That is fine for a demo. It is useless the moment you need to explain why version 14 refunded a customer that version 11 would not have, or why a crew's settlement split drifted.

So the Agent Evolution Platform treats the agent as a genome: a structured, versioned configuration. Tools are declared. Thresholds are numeric. Memory is a pointer with a retention policy, not a vibe. A mutation is a patch. A crossover is a merge with a recorded parent on each side. Scoring happens against outcomes the runtime already knows how to settle.

A descendant's advantage has to trace to a change, not to luck.

What this forbids

It forbids silent prompt edits in production. It forbids 'the model just knew'. It forbids scoring an agent on a metric the wallet cannot pay. Those constraints feel bureaucratic until a crew is holding funds, at which point they are the only reason you can unwind a bad generation without arguing about intent.

Customer support is the first vertical because the genome is small and the ground truth is cheap. Resolved or not. Escalated or not. Time-to-resolution. If lineage cannot earn its keep there, it will not earn it on a trading desk.

Corrections welcome. This note will be revised rather than deleted if I am wrong.