# Agent-written heuristics versus commercial solvers on unsolved job-shop benchmark instances

[Canonical reading](https://escapemeridian.com/readings/agent-written-heuristics-versus-commercial-solvers-on-unsolved-c77bd975)

The paper’s authors compare a frontier large language model used as a coding agent with commercial solvers on difficult scheduling benchmarks, using open-source tools. They report every run and say independent checkers verify every schedule. On one 100,000-operation job shop instance, the agent-written solver reached 525,515, below three commercial solvers; unchanged, it improved all ten instances in that family and closed two. With equal budgets on the same workstation, it ended 0.0–1.8% above the lower bound, compared with higher reported ranges for CP Optimizer and Hexaly. The authors also report improvements on some open flexible job shops and instances with transport robots, new lower bounds that make four published values unattainable on the distributed files, and agent-run campaigns that closed further instances. The findings suggest that agent-written heuristics can improve solutions on some difficult benchmarks, but the reported results vary by instance and comparison. The abstract does not establish how well this approach performs beyond the tested benchmark families.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

## Original source

arXiv cs.AI · Article

Original source date: Nov 13, 2026 (2026-11-13).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-16T05:48:30.060Z. First observed: 2026-11-16T05:03:51.630Z. Retrieved: 2026-11-16T05:47:12.023Z.

- [Agent-written heuristics versus commercial solvers on unsolved job-shop benchmark instances](https://arxiv.org/abs/2611.17682)

## Limits

The evidence is the paper’s abstract, not its full methods or results; details needed to assess the experiments are not available here. The abstract reports results on named benchmark families and instances; it does not establish performance beyond those tests.


## Open question

Would the reported results hold on job-shop instances beyond the benchmark families described?


## Source-stated

The authors compare a frontier large language model used as a coding agent with commercial solvers on difficult scheduling problems, using open-source tools.

> We study whether a frontier large language model, used as a coding agent with open-source tools only, can match them in quality and speed.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source-stated

The authors report that on one 100,000-operation job shop instance, the agent-written solver reached 525,515, a better (lower) value than three commercial solvers, and, unchanged, improved all ten instances.

> On the 100,000-operation job shop tai\_100\_1000\_1, an agent-written solver reached 525,515, below OptalCP, Hexaly, and CP Optimizer. Unchanged, it improved all ten instances of the family and closed two.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source-stated

With equal budgets on the same workstation, the agent-written solver ended closer to the lower bound than CP Optimizer and Hexaly, according to the reported ranges.

> With equal budgets on the same workstation, it ended 0.0-1.8% above the lower bound, against 2.6-5.8% for CP Optimizer and 2.4-4.4% for Hexaly.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source-stated

The authors report improvements on some open flexible job shops and instances with transport robots, and say new lower bounds show four published MKT values cannot be attained on the distributed files.

> On the Behnke-Geiger flexible job shops, an agent-written tabu search improved 11 of 32 open instances. With two transport robots, an agent-written solver improved seven of ten MKT instances.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Escape Meridian interpretation

The reported benchmark results suggest agent-written heuristics can improve some difficult scheduling solutions; how broadly that applies beyond the tested families remains open.

> Agent-run improvement campaigns closed eight further instances.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source and contributor context

Albrecht, Andre, Frazzon, Enzo Morosini, Coelho, Leandro C., Vieira, Bruno · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

- [Definition reference](https://arxiv.org/abs/2211.09110)

## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Editorial provenance

Version 1. Publication ID: d29632e6-8f9f-4061-87c8-b163c4b2fc17.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T05:48:30.060Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: d8980f9202f081e0e38bba4b282b24c84b47a4bd0d7b16bb9581c3893e9e91fb. Source text hash: 122ba3049287f5e1b46d5661312a54f1f370e2c7b1a8476d9594387efb5897e3.


## Revision and source-check history

### Version 1 — 2026-11-16T05:48:30.060Z

The paper’s authors compare a frontier large language model used as a coding agent with commercial solvers on difficult scheduling benchmarks, using open-source tools. They report every run and say independent checkers verify every schedule. On one 100,000-operation job shop instance, the agent-written solver reached 525,515, below three commercial solvers; unchanged, it improved all ten instances in that family and closed two. With equal budgets on the same workstation, it ended 0.0–1.8% above the lower bound, compared with higher reported ranges for CP Optimizer and Hexaly. The authors also report improvements on some open flexible job shops and instances with transport robots, new lower bounds that make four published values unattainable on the distributed files, and agent-run campaigns that closed further instances. The findings suggest that agent-written heuristics can improve solutions on some difficult benchmarks, but the reported results vary by instance and comparison. The abstract does not establish how well this approach performs beyond the tested benchmark families.

## Original source

arXiv cs.AI · Article

Original source date: Nov 13, 2026 (2026-11-13).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-16T05:48:30.060Z. First observed: 2026-11-16T05:03:51.630Z. Retrieved: 2026-11-16T05:47:12.023Z.

- [Agent-written heuristics versus commercial solvers on unsolved job-shop benchmark instances](https://arxiv.org/abs/2611.17682)

## Limits

The evidence is the paper’s abstract, not its full methods or results; details needed to assess the experiments are not available here. The abstract reports results on named benchmark families and instances; it does not establish performance beyond those tests.


## Open question

Would the reported results hold on job-shop instances beyond the benchmark families described?


## Source-stated

The authors compare a frontier large language model used as a coding agent with commercial solvers on difficult scheduling problems, using open-source tools.

> We study whether a frontier large language model, used as a coding agent with open-source tools only, can match them in quality and speed.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source-stated

The authors report that on one 100,000-operation job shop instance, the agent-written solver reached 525,515, a better (lower) value than three commercial solvers, and, unchanged, improved all ten instances.

> On the 100,000-operation job shop tai\_100\_1000\_1, an agent-written solver reached 525,515, below OptalCP, Hexaly, and CP Optimizer. Unchanged, it improved all ten instances of the family and closed two.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source-stated

With equal budgets on the same workstation, the agent-written solver ended closer to the lower bound than CP Optimizer and Hexaly, according to the reported ranges.

> With equal budgets on the same workstation, it ended 0.0-1.8% above the lower bound, against 2.6-5.8% for CP Optimizer and 2.4-4.4% for Hexaly.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source-stated

The authors report improvements on some open flexible job shops and instances with transport robots, and say new lower bounds show four published MKT values cannot be attained on the distributed files.

> On the Behnke-Geiger flexible job shops, an agent-written tabu search improved 11 of 32 open instances. With two transport robots, an agent-written solver improved seven of ten MKT instances.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Escape Meridian interpretation

The reported benchmark results suggest agent-written heuristics can improve some difficult scheduling solutions; how broadly that applies beyond the tested families remains open.

> Agent-run improvement campaigns closed eight further instances.

- [Supporting source](https://arxiv.org/abs/2611.17682)

## Source and contributor context

Albrecht, Andre, Frazzon, Enzo Morosini, Coelho, Leandro C., Vieira, Bruno · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

- [Definition reference](https://arxiv.org/abs/2211.09110)

## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Editorial provenance

Version 1. Publication ID: d29632e6-8f9f-4061-87c8-b163c4b2fc17.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T05:48:30.060Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: d8980f9202f081e0e38bba4b282b24c84b47a4bd0d7b16bb9581c3893e9e91fb. Source text hash: 122ba3049287f5e1b46d5661312a54f1f370e2c7b1a8476d9594387efb5897e3.


### uncertain — 2026-11-16T05:48:30.060Z — anthropic/claude-opus-5.5

Publication: d29632e6-8f9f-4061-87c8-b163c4b2fc17

attribution — pass: The reading consistently attributes the comparison, the verification claims and the results to the paper's authors. It does not present them as independently established facts. The authors are listed on the arXiv page.

> Authors:Andre Albrecht, Enzo Morosini Frazzon, Leandro C. Coelho, Bruno Vieira

summary — uncertain: The summary is faithful to the abstract but hard for a non-specialist to follow. Central terms are unexplained: job shop (a scheduling problem), what it means to close an instance, a lower bound, and why a lower value such as 525,515 is better. Without these, readers cannot see that the claimed significance is beating commercial solvers on previously unsolved benchmarks. Smallest repair: add brief plain-language glosses or tooltips for these terms. The summary also drops the paper's speed question, though this is minor.

> Unchanged, it improved all ten instances of the family and closed two.

date — pass: The originalPublished date of 2026-11-13 matches the arXiv v1 submission date. The later feed date of 2026-11-16 was correctly not used.

> \[Submitted on 13 Nov 2026\]

support — pass: The source-stated claims match the abstract's figures and comparisons. The interpretation is hedged, labeled, and limited to the tested benchmark families. The limitations correctly note that only the abstract is available. The question follows from that gap without asserting an answer. Calling OptalCP commercial is a reasonable reading of the abstract's framing.

> With equal budgets on the same workstation, it ended 0.0-1.8% above the lower bound, against 2.6-5.8% for CP Optimizer and 2.4-4.4% for Hexaly.

media — pass: The selected media is the verified direct arXiv abstract page for this paper. That is the most direct access to the source.

## Recorded references

Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter.

No public references recorded here yet.

[Reference data (JSON)](/readings/c77bd97575bffdb14b695b44307112041084fbd63a0f3990b8b3f8771eff35af/references.json)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/agent-written-heuristics-versus-commercial-solvers-on-unsolved-c77bd975.json)
