🤖 My AI Agents Version Themselves: How We Built Self-Evolving Personas Using Semantic Versioning
agents ai-safety
| Source: Dev.to | Original article
A research team at Oslo‑based startup Cognition Labs has released a prototype in which autonomous AI agents rewrite their own code, tag each iteration with a semantic version number and store the changes on disk. The agents are given a single hard rule – never repeat a mistake – and are allowed to experiment, fail and learn without human intervention. Within days the system produced a hierarchy of “personas” that each carried a version label such as 1.2.3, documenting functional upgrades, bug fixes and newly added capabilities. The versioning scheme mirrors software‑development practices, enabling the team to track progress, roll back regressions and audit the evolution of each agent.
The breakthrough matters because it moves self‑improvement from a theoretical concept to a concrete engineering workflow. By embedding version control directly into the agent’s runtime, developers can monitor emergent behaviour, enforce safety constraints and maintain reproducibility – a long‑standing hurdle for open‑ended AI. The approach also dovetails with recent work on self‑monitoring multi‑timescale agents, which we covered on 16 April 2026, showing that metacognitive loops can be harnessed for continuous learning. If agents can reliably avoid past errors while iterating autonomously, the cost of fine‑tuning large language models could drop dramatically, opening the door to personalised assistants that evolve with individual users or domain‑specific tasks.
What to watch next is the rollout of the framework beyond the lab. Cognition Labs plans an open‑source SDK later this quarter, inviting developers to embed the versioning engine in chatbots, robotics and enterprise automation. Regulators are already asking how such self‑modifying systems will be audited, and the EU’s AI Act may need to address version‑controlled agents explicitly. The next few months will reveal whether semantic versioning can become a standard safety net for the next generation of self‑evolving AI.
Sources
Back to AIPULSEN