Agent-written solver beats commercial tools on open job-shop benchmark
An arXiv preprint reports that a large language model used as a coding agent, restricted to open-source tools, wrote schedulers that outperformed commercial solvers on several unsolved benchmark instances.
In plain words
Job-shop scheduling means finding the best order for thousands of tasks on shared machines, a classic hard optimization problem usually handled by commercial software. The authors had a frontier large language model act as a coding agent that wrote its own solvers using only open-source tools. On some long-unsolved benchmark cases, these agent-written solvers found better schedules than established commercial products.
What was announced
The preprint's central test is the 100,000-operation job shop tai_100_1000_1, where an agent-written solver reached 525,515, a result the authors say falls below OptalCP, Hexaly, and CP Optimizer. With equal time budgets on the same workstation, the agent-written solver finished 0.0-1.8% above the lower bound, while CP Optimizer ranged 2.6-5.8% and Hexaly 2.4-4.4%. The authors state that every run is reported and that independent checkers verified every schedule. Unchanged, the solver improved all ten instances in the family and closed two, meaning it matched the best known results on them.
The paper extends the approach to other problem families. On the Behnke-Geiger flexible job shops, an agent-written tabu search improved 11 of 32 open instances, and with two transport robots an agent-written solver improved seven of ten MKT instances. The authors also report new lower bounds showing that four published MKT values cannot be attained on the distributed instance files, and that agent-run improvement campaigns closed eight further instances. The abstract does not name the language model used, describe how the agent was prompted, or state how much human intervention the coding process required.
What it does not show
- The evidence is the paper's abstract only; the full article's methods, model identity, and experimental details are not covered.
- Results are self-reported by the authors with their own verification setup; no independent confirmation is included in the excerpt.
- The abstract does not specify which frontier language model was used or the cost and effort of running the agent.
For security and IT teams
- Scope
- The reported gains apply to specific job-shop benchmark families (tai, Behnke-Geiger, MKT) under the paper's equal-budget workstation setup, not to scheduling software generally.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.