SEVerA: Self-Evolving Agents That Can Formally Prove They're Correct — A New Framework Closes the Safety Gap in Autonomous Code Rewriting
Most conversations about agent safety focus on capability restrictions — don't let agents delete files, don't let them send emails without approval. SEVerA (Self-Evolving Verified Agents) targets a harder and less discussed problem: when an agent autonomously rewrites its own programs, what guarantees do you have that the evolved behavior still satisfies correctness requirements? The answer from existing frameworks is: none. SEVerA introduces formal guarantees into the self-evolution loop through a three-stage framework of synthesis, verification, and learning.
The key technical contribution is Formally Guarded Generative Models (FGGM), which wrap each sub-model call in a rejection sampler with a verified fallback. Every output from every model call is guaranteed to satisfy a formally specified contract — expressed in first-order logic — regardless of input. This separates two concerns that self-evolving systems typically conflate: correctness (verified statically, guaranteed to hold) and performance (improved by gradient-based learning, free to optimize within the verified constraints). Evaluations covered Dafny program verification, symbolic math synthesis, and policy-compliant agentic tool use.
For teams deploying agents in high-stakes environments — finance, healthcare, infrastructure — SEVerA addresses the gap between "we tested it empirically" and "it's formally correct." As agents become capable enough to rewrite their own operational code, the distinction between those two levels of confidence becomes a first-class engineering concern rather than an academic one. The FGGM abstraction is directly applicable to any system where hard output constraints need to hold across autonomous agent iterations.