Autogenesis: A Self-Evolving Agent Protocol

We introduce Autogenesis, a protocol that enables LLM-based agents to autonomously evolve themselves — rewriting their own prompts, tools, and memory at runtime without human intervention.

Why Agent Self-Evolution?

Current LLM agent systems are fundamentally static. Their prompts are written once by engineers. Their tools are fixed at deployment. Their behaviors do not improve from experience. This is a severe limitation: real-world tasks are diverse and unpredictable, and static agents plateau quickly as task complexity grows.

Autogenesis asks: can an agent monitor its own failures, diagnose what went wrong, and rewrite itself to do better — autonomously?

Architecture

Autogenesis is built around two tightly coupled layers:

1. Resource Substrate Protocol Layer

Everything in an Autogenesis agent — prompts, tools, memory, even sub-agents — is modeled as a versioned resource with an explicit lifecycle:

• Create → Activate → Suspend → Terminate
• Every mutation produces a new version; prior versions are retained for rollback
• Resources expose typed interfaces, enabling safe composition and substitution

This versioning foundation is what makes self-modification safe: the agent can experiment with changes and revert if performance degrades.

2. Self-Evolution Protocol Layer

The evolution layer operates as a closed feedback loop:

1. Monitor — tracks task outcomes, tool call success rates, and reasoning quality
2. Diagnose — identifies failure patterns and root causes across recent episodes
3. Propose — generates candidate resource mutations (prompt rewrites, new tools, memory restructuring)
4. Evaluate — tests candidates against a held-out task distribution
5. Commit — promotes successful mutations; discards or rolls back failures

This loop runs continuously during deployment, not just at training time.

Results

Autogenesis demonstrates consistent improvement on benchmarks involving complex planning and multi-step tool use. Key findings:

• Agents improve measurably across repeated task episodes without any human-written updates
• The versioned resource model prevents catastrophic self-modification — rollback is triggered automatically when performance regresses
• Self-evolved prompts outperform hand-crafted prompts on out-of-distribution tasks

Connection to Cradle

Autogenesis shares a philosophical foundation with Cradle: both systems treat self-improvement as a first-class capability, not an afterthought. Where Cradle builds a skill library through interaction, Autogenesis goes further — evolving the agent’s core resources at the protocol level.

Together, they represent a research direction toward agents that genuinely improve through experience.

Links

• Paper: arXiv:2604.15034
• GitHub: DVampire/Autogenesis