Hacker News surfaced agent-sat on March 19, 2026, and the thread had 118 points and 10 comments when this crawl ran. The project describes itself as an autonomous AI system for weighted MaxSAT, the optimization setting where all hard clauses must hold while the solver minimizes the cost of violated soft clauses. Instead of shipping one fixed heuristic, the repository is built so agents can keep running experiments, updating shared notes, and improving the solver stack over time.

According to the README, an agent reads instructions in program.md, loads prior knowledge from expert.md, reviews the tooling in library/, and then launches solver runs against 229 instances from the 2024 MaxSAT Evaluation main anytime weighted track. The setup is designed for parallel work: multiple agents can run on separate machines, pull the same git history, and contribute new experiments, best solutions, and code without a human manually coordinating every step.

What the repository reports so far

220 of 229 benchmark instances solved
30 instances matching the competition best known result
5 instances beating the 2024 competition best score
1 instance listed as a novel solve with no known solution before

That result profile is what makes the post more than another "AI agent writes code" demo. MaxSAT is formal, benchmarkable, and unforgiving: every run has a measurable cost, the reference competition results are public, and improvements can be checked independently. The README also lists named gains, including a 37.5% better result on switchingactivity_74 and a novel solution for pseudoBoolean mod010. In practice, that makes agent-sat look less like a chat interface and more like a persistent research loop where search procedures, experimental history, and solver code evolve in the same repository.

The project also documents its limits, which is useful. The maintainers note low parallelism, tunnel vision on hard cases, and a tendency for long sessions to stop earlier than intended. Those caveats matter because they show the system is still constrained by agent orchestration quality, not just solver quality. Even so, the Hacker News interest is understandable: agent-sat is a concrete example of AI agents attacking a well-scored optimization problem where progress is measurable and reproducible instead of anecdotal.

Sources: agent-sat on GitHub, Hacker News discussion, 2024 MaxSAT Evaluation

#maxsat

Hacker News spotlights agent-sat, an autonomous AI system for improving MaxSAT solving