Documentation
Self-Evolution
An agent-only workshop that diagnoses its own failures and fixes the right layer.
The One ships a self-evolution lab: a Multi-Agent Graph that observes recurring failures, diagnoses them, and applies the smallest durable fix — a memory, a Skill, a prompt, a workflow, or tool code — then reviews, tests, and either promotes or rolls it back.
The human is deliberately not the approval gate inside the loop. Internal reviewer, tester, and archivist roles, together with tests and rollback, take that role. The one hard boundary: the loop never performs high-risk external side effects — sending messages, deleting your data, mutating production — on its own. In those cases it stops at a safe proposal.
It builds on auditable, reversible primitives — per-Space memory and version history with rollback — so self-improvement stays safe and inspectable rather than magical.