An Organization Chart of LLM Agents in a One-Person Software Company: An Empirical Case Study of Routing, User-Rule Memory, and Rework

Multi-agent frameworks borrow the vocabulary of companies: roles, managers, hand-offs. Published reports of sustained real use of such structures appear to be few. We report a case study of a one-person software company that used an organization of LLM subagents in Claude Code alongside its main sessions, grouped into 15 departments with a chief agent each and 126 registered agents at the end of the observation window. Using only aggregates computed from local session logs, a host-application database, git history, and the owner's rule memory, we describe how the organization grew (from 33 registered agents in early October to 126 in mid-November), how 81 rule-memory files accumulated and were revised, and how often the owner asked for rework after agent output. A first labeling pass with a local LLM agreed poorly with the owner's judgments (binary κ = 0.05 on 30 items) and was discarded. A revised labeler reached binary κ = 0.526 on a 15-item test split made after error analysis of all 30 items, and 0.625 when fixed card-button replies are excluded (n = 12, 95% CI −0.09 to 1.00). With this provisional labeler, 10.4% of the owner's typed follow-up messages were labeled as asking for a correction or a redo (95% CI 7.3–13.8). The ordering of this rate across three organizational periods reversed between labeler versions, so we report the rate as labeler-sensitive and make no claim about the effect of restructuring. The labeler marked none of 44 judgeable delegations as a request to change the assigned agent, but the reference set held no such case, so routing accuracy could not be measured. We map failure types to the MAST taxonomy, give anonymized failure cases and a local media-generation pipeline, and document the filters used to exclude client work and personal data from the study.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24212965
Primary Topic
Software Engineering Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

An Organization Chart of LLM Agents in a One-Person Software Company: An Empirical Case Study of Routing, User-Rule Memory, and Rework

Jaein Lee
Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
preprint

An Organization Chart of LLM Agents in a One-Person Software Company: An Empirical Case Study of Routing, User-Rule Memory, and Rework

Jaein Lee
preprint en

Abstract

Multi-agent frameworks borrow the vocabulary of companies: roles, managers, hand-offs. Published reports of sustained real use of such structures appear to be few. We report a case study of a one-person software company that used an organization of LLM subagents in Claude Code alongside its main sessions, grouped into 15 departments with a chief agent each and 126 registered agents at the end of the observation window. Using only aggregates computed from local session logs, a host-application database, git history, and the owner's rule memory, we describe how the organization grew (from 33 registered agents in early October to 126 in mid-November), how 81 rule-memory files accumulated and were revised, and how often the owner asked for rework after agent output. A first labeling pass with a local LLM agreed poorly with the owner's judgments (binary κ = 0.05 on 30 items) and was discarded. A revised labeler reached binary κ = 0.526 on a 15-item test split made after error analysis of all 30 items, and 0.625 when fixed card-button replies are excluded (n = 12, 95% CI −0.09 to 1.00). With this provisional labeler, 10.4% of the owner's typed follow-up messages were labeled as asking for a correction or a redo (95% CI 7.3–13.8). The ordering of this rate across three organizational periods reversed between labeler versions, so we report the rate as labeler-sensitive and make no claim about the effect of restructuring. The labeler marked none of 44 judgeable delegations as a request to change the assigned agent, but the reference set held no such case, so routing accuracy could not be measured. We map failure types to the MAST taxonomy, give anonymized failure cases and a local media-generation pipeline, and document the filters used to exclude client work and personal data from the study.

Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon