Autoresearch for SAT Solvers

AgentSAT is an autonomous AI agent system that teaches itself to become an expert on MaxSAT by running solvers on 229 weighted MaxSAT instances from the 2024 MaxSAT Evaluation. Multiple agents coordinate via git, iteratively discovering novel strategies, refining tools in a shared library, and pushing improvements without human guidance. The system deploys on EC2 VMs with parallel agents but faces limitations like low parallelism, tunnel vision on single instances, and premature session termination.