Agents Beat Commercial Solvers, and Both Papers Publish the Failed Runs
Commercial solvers like CP Optimizer and Hexaly are what you buy when scheduling gets hard. A paper in Monday's arXiv batch had a frontier model, working as a coding agent with open-source tools only, write heuristics against them. It won. Paper: https://arxiv.org/abs/2611.17682
On tai_100_1000_1, a job shop with 100,000 operations, the agent-written solver reached 525,515, below OptalCP, Hexaly and CP Optimizer. Unchanged, it improved all ten instances of that family and closed two. With equal budgets on the same workstation it finished 0.0 to 1.8% above the lower bound, against 2.6 to 5.8% for CP Optimizer and 2.4 to 4.4% for Hexaly. An agent-written tabu search improved 11 of 32 open flexible job-shop instances, and new lower bounds showed four published values on one benchmark set cannot actually be attained.
A second paper did the same thing in physics. The Thomson problem asks how to place N charges on a sphere at minimum energy, and the best known answers sit in the Cambridge Cluster Database. A search pipeline built and run with Claude, through chat and Claude Code, found configurations for N = 912, 1382 and 2502 below the database values, certified by exact integer arithmetic. Paper: https://arxiv.org/abs/2611.17776
What makes both worth reading is not the records. It is the bookkeeping. The scheduling paper reports every run and has independent checkers verify every schedule. The Thomson paper reports every search including the failures, and admits that polishing alone produced two spurious records until candidates were compared against polished database energies. Its first rule for others: verify candidates with code separate from the search code. That is the verifier-is-the-product thread in its most practical form. An agent can search all night. The part worth paying for is the checker it does not get to write.
← Back to all articles
On tai_100_1000_1, a job shop with 100,000 operations, the agent-written solver reached 525,515, below OptalCP, Hexaly and CP Optimizer. Unchanged, it improved all ten instances of that family and closed two. With equal budgets on the same workstation it finished 0.0 to 1.8% above the lower bound, against 2.6 to 5.8% for CP Optimizer and 2.4 to 4.4% for Hexaly. An agent-written tabu search improved 11 of 32 open flexible job-shop instances, and new lower bounds showed four published values on one benchmark set cannot actually be attained.
A second paper did the same thing in physics. The Thomson problem asks how to place N charges on a sphere at minimum energy, and the best known answers sit in the Cambridge Cluster Database. A search pipeline built and run with Claude, through chat and Claude Code, found configurations for N = 912, 1382 and 2502 below the database values, certified by exact integer arithmetic. Paper: https://arxiv.org/abs/2611.17776
What makes both worth reading is not the records. It is the bookkeeping. The scheduling paper reports every run and has independent checkers verify every schedule. The Thomson paper reports every search including the failures, and admits that polishing alone produced two spurious records until candidates were compared against polished database energies. Its first rule for others: verify candidates with code separate from the search code. That is the verifier-is-the-product thread in its most practical form. An agent can search all night. The part worth paying for is the checker it does not get to write.
Comments