Intel / article

MAGS: Multi-agent Auto-formalization Guarantees Safety for Agentic Outputs

Source: arxiv.org
Author: Albert Wu, Nicholas Roberts, Tzu-Heng Huang, Haoran Lin, Gil Friedman, Sungjun Cho, Gabriel Orlanski, Frederic Sala
Date: Published Sep 18, 2026

Context

MAGS applies machine-checkable safety requirements to AI-generated programs. It fixes human-audited requirements in advance, translates generated code into Dafny, a language supporting formal verification, and uses verification feedback to repair violations before producing executable code. This makes the stated requirements explicit enough for a software verifier to check. Research abstract.

The authors report programs with verified properties for all 220 evaluated examples. That guarantee depends on what the formal specification captures: separate safety and functional evaluations still found failures where the formalized behavior did not fully represent the intended task. The result makes accurate specification a central part of safety, alongside proving that code satisfies it. Results and qualification.

Key takeaways · AI-distilled

MAGS routes generated code through Dafny, a machine-checkable verification language, repairing any safety violations using verifier feedback before compiling back to executable code.
Across 220 test cases (100 CUDA kernels, 100 terminal scripts, 20 robotic-arm tasks), MAGS achieved a 100% success rate producing programs with formally verified safety guarantees against frozen specifications.
Independent safety and functional evaluations still found failures in cases where the auto-generated formal (Dafny) version didn't fully capture the intended real-world behavior -- formal verification only guarantees what it was told to check.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source arxiv.org