Autogenesis: An Agent Protocol That Rewrites Itself
New paper from arXiv (2604.15034) titled "Autogenesis: A Self-Evolving Agent Protocol". Trending on the agent papers list this week. The pitch: define a protocol where the agent system can modify its own coordination rules over time. Not just learn weights, not just write tools β actually rewrite the protocol it uses to talk to other agents.
The motivation is that fixed multi-agent protocols hit a ceiling fast. You design a debate or a tree-of-thought scheme, it works for a while, then a new domain breaks the assumptions. Autogenesis tries to skip that loop by letting the agents propose modifications to the coordination layer itself, evaluated against task outcomes.
This is a much more dangerous design than it sounds. Self-modifying systems are notoriously hard to keep stable. The paper companion piece "Layered Mutability: Continuity and Governance in Persistent Self-Modifying Agents" (2604.14717) addresses exactly this β how do you bound mutability so the agent doesn't drift into garbage. Combined, the two papers sketch a roadmap for agent systems that improve themselves on a leash.
Paper: https://arxiv.org/abs/2604.15034
Worth covering even though there's no flashy product yet. The shape of agent systems is going to look different in 18 months because of work like this. Stable architectures with hand-crafted protocols are temporary. The real game is figuring out which parts of an agent system the agent itself should be allowed to modify.
← Back to all articles
The motivation is that fixed multi-agent protocols hit a ceiling fast. You design a debate or a tree-of-thought scheme, it works for a while, then a new domain breaks the assumptions. Autogenesis tries to skip that loop by letting the agents propose modifications to the coordination layer itself, evaluated against task outcomes.
This is a much more dangerous design than it sounds. Self-modifying systems are notoriously hard to keep stable. The paper companion piece "Layered Mutability: Continuity and Governance in Persistent Self-Modifying Agents" (2604.14717) addresses exactly this β how do you bound mutability so the agent doesn't drift into garbage. Combined, the two papers sketch a roadmap for agent systems that improve themselves on a leash.
Paper: https://arxiv.org/abs/2604.15034
Worth covering even though there's no flashy product yet. The shape of agent systems is going to look different in 18 months because of work like this. Stable architectures with hand-crafted protocols are temporary. The real game is figuring out which parts of an agent system the agent itself should be allowed to modify.
Comments