A self-directed AI agent has achieved remarkable results in the 2024 MaxSAT Evaluation, solving 220 out of 229 weighted MaxSAT instances without any human guidance. The autonomous system, built using Claude Code, iteratively improves its strategies by running solvers, analyzing results, and updating its knowledge base. It discovered novel approaches that outperformed the competition in several instances.

Working across multiple virtual machines, the agent communicates through git, pulling the latest solutions and pushing improvements. Starting with 229 benchmark problems from the 2024 competition, it found optimal solutions for 30 instances and achieved better results than the competition in five cases. The agent also solved one problem with no previously known solution, demonstrating genuine discovery capabilities.

Key breakthroughs include a 637.5% improvement on switchingactivity_74, 327.8% on synplicate dag_run2_10_size_11, and 101.1% on switchingactivity_688. The agent developed techniques like core-guided search, biased-SAT for breaking local optima, and alternating CWLS with WalkSAT. Despite limitations including low parallelism and occasional tunnel vision on single problems, the system continues to evolve its library of solver tools and strategies, maintaining a living knowledge base in expert.md.