Project Name
EVA
Repository URL
https://github.com/arturkorb3/eva-evolutional-agent
Category
Agent Evolution and Self-Improvement
Description
EVA is a research prototype for controlled self-evolving LLM agents. Instead of providing one fixed agent architecture, EVA provides a small immutable kernel and a minimal seed organism whose agent runtime can evolve over time.
Changes do not happen live. EVA must fork a candidate release, modify that candidate, pass tests, supervisor/kernel gates, prompt-surface admission checks, and promotion checks before the candidate becomes active. Every promotion is tracked in a rollback ledger.
Why should it be included?
EVA explores a safety-first but architecture-agnostic approach to agent self-evolution: the agent is allowed to evolve its own runtime — tools, loop, memory, prompts, self-model, and gates — but only through gated candidate releases.
The core idea is that agent architecture should grow from repeated friction and real user needs instead of being fully designed up front. Different forks can therefore grow different 'organs' under different tasks and constraints, while still being bounded by an immutable kernel, tests, promotion gates, prompt-surface auditing, and rollback.
It is a very young project and currently below the usual star threshold, so I understand if it does not yet meet the traction bar. However, I think the combination of immutable kernel, evolvable release state, friction-driven growth, gated promotion, prompt-surface safety checks, and rollback ledger may be novel/unique enough as an approach to controlled self-modification to be worth a look.
Open to feedback either way, and happy to resubmit once it has more traction.
Related Paper (optional)
N/A
Project Name
EVA
Repository URL
https://github.com/arturkorb3/eva-evolutional-agent
Category
Agent Evolution and Self-Improvement
Description
EVA is a research prototype for controlled self-evolving LLM agents. Instead of providing one fixed agent architecture, EVA provides a small immutable kernel and a minimal seed organism whose agent runtime can evolve over time.
Changes do not happen live. EVA must fork a candidate release, modify that candidate, pass tests, supervisor/kernel gates, prompt-surface admission checks, and promotion checks before the candidate becomes active. Every promotion is tracked in a rollback ledger.
Why should it be included?
EVA explores a safety-first but architecture-agnostic approach to agent self-evolution: the agent is allowed to evolve its own runtime — tools, loop, memory, prompts, self-model, and gates — but only through gated candidate releases.
The core idea is that agent architecture should grow from repeated friction and real user needs instead of being fully designed up front. Different forks can therefore grow different 'organs' under different tasks and constraints, while still being bounded by an immutable kernel, tests, promotion gates, prompt-surface auditing, and rollback.
It is a very young project and currently below the usual star threshold, so I understand if it does not yet meet the traction bar. However, I think the combination of immutable kernel, evolvable release state, friction-driven growth, gated promotion, prompt-surface safety checks, and rollback ledger may be novel/unique enough as an approach to controlled self-modification to be worth a look.
Open to feedback either way, and happy to resubmit once it has more traction.
Related Paper (optional)
N/A