ai4he/supervisory-capacity-model

Supervisory Capacity Model

Dataset accompanying the paper How Many Agents Can One Supervisor Track? A Mechanistic Capacity Model and Measurement Protocol for LLM-Agent Teams (Human-AI Coevolution Workshop (HAIC) at NeurIPS 2026).

Parameters and outputs of the mechanistic supervisory-capacity model (capacity parameters, service-time variance, simulation curves and summary), the computational probe pilot (per-row results and summary), and the scripts that produce them.

Paper	How Many Agents Can One Supervisor Track? A Mechanistic Capacity Model and Measurement Protocol for LLM-Agent Teams
Venue	Human-AI Coevolution Workshop (HAIC) at NeurIPS 2026
Authors	Carlos Toxtli, Manuel Delaflor
License	Creative Commons Attribution 4.0 International (cc-by-4.0); code: MIT License
Dataset DOI	10.5281/zenodo.24179030
Version	1.0.0
Maintainer	Human-AI Empowerment Lab, Clemson University

Paper abstract

Agentic deployments can assign many parallel agents to one supervisor without specifying when that supervisor can no longer recover the team's state. Human-factors research already establishes relevant constructs such as situation awareness, span of control, neglect tolerance, and multiple-resource workload, but language-model teams distribute state across text and can fail silently. This position paper translates those constructs into a measurement protocol designed for later calibration to people rather than a new human experiment. Typed event rates, isolated service costs, and a throughput budget define a nominal load-balance index; the midpoint of a separately fitted capacity curve remains an empirical quantity. Synthetic controls recover a known capacity boundary, while a seeded queue shows that deadline pressure can separate the fitted midpoint from nominal load balance. The resulting event ontology, operating envelope, executable stress tests, and preregisterable falsification criteria allow system designers to test staffing or interface interventions within a declared team-monitor channel. They do not turn synthetic behavior into evidence about human or model supervisors.

Dataset structure

Path	Format	Size	Records
code/capacity_parameters.py	py	5.6 KB
code/capacity_simulation.py	py	9.9 KB
code/make_parameter_table.py	py	1.0 KB
code/probe_pilot.py	py	10.5 KB
code/service_variance.py	py	3.2 KB
data/capacity_parameters.json	json	3.5 KB
data/probe_pilot_results.json	json	31.5 KB
data/probe_pilot_rows.jsonl	jsonl	20.6 KB	120
data/service_variance.json	json	828 B
data/simulation_curves.dat	dat	281 B
data/simulation_summary.json	json	5.1 KB
docs/SUPPLEMENT_README.md	md	1.5 KB

12 files, 93.4 KB in total (metadata files excluded). Records are rows for CSV/TSV/JSONL files and list length for JSON arrays.

Loading the data

from datasets import load_dataset

ds = load_dataset("ai4he/supervisory-capacity-model", "probe_pilot_rows")

Available subsets: probe_pilot_rows. All other files can be downloaded with huggingface_hub.snapshot_download or from the Zenodo record.

Known limitations

• Simulation and computational-pilot data only; the human study the protocol specifies has not been run.

Personal and sensitive information

None.

License

The data and documentation are released under the Creative Commons Attribution 4.0 International. Source code files (.ipynb, .js, .m, .py, .sh) are released under the MIT License.

Citation

If you use this dataset, please cite the paper:

@inproceedings{Toxtli2026Supervisor,
 title = {How Many Agents Can One Supervisor Track? A Mechanistic Capacity Model and Measurement Protocol for LLM-Agent Teams},
 author = {Toxtli, Carlos and Delaflor, Manuel},
 booktitle = {Human-AI Coevolution: Measuring Human-Agent Teams in the Agentic Era (HAIC) Workshop at NeurIPS 2026},
 address = {Atlanta, GA},
 year = {2026},
 month = dec,
 note = {Workshop paper},
 url = {https://openreview.net/forum?id=czINRnBTpf}
}

and the dataset:

@dataset{Toxtli2026SupervisoryCapacityModelData,
  author = {Toxtli, Carlos and Delaflor, Manuel},
  title = {{Supervisory Capacity Model}: Dataset for ``How Many Agents Can One Supervisor Track? A Mechanistic Capacity Model and Measurement Protocol for LLM-Agent Teams''},
  year = {2026},
  version = {1.0.0},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.24179030},
  url = {https://doi.org/10.5281/zenodo.24179030}
}

Further documentation: docs/SUPPLEMENT_README.md.