Coase-Information Theory for AI Orgs

▶ Live interactive companion → https://galatheus-labs.github.io/coase-info-theory/ ·

Paper (PDF)·

Paper (TeX)·

Galatheus

This repo packages a small set of static web apps plus formal paper material for the Coase–Information Theory project.

The central claim is:

Firms are a special case of agent-boundary formation. Agents coalesce when shared representation reduces surprise and action loss more than it raises coordination cost, and split when modularity lowers cost more than it raises interface surprise.

The practical implication:

Organizations are information architectures. Their economic performance depends on how efficiently they convert distributed, noisy signals into coordinated action under latency, cost, and distortion constraints.

The browser demo translates that claim into tangible software.

The companion page opens with a 60-second animated walkthrough of the thesis — a firm is an agent in a market, and is itself made of agents — then lets you drive the model across concrete "current problems":

• monolith vs microservices (the granularity sweet spot, m* = √(A/B)

), - managed database / vendor dependency,

• support triage agent,
• coding maintenance agent,
• security alert triage,
• procurement and AP automation.

A boundary phase map shows which architecture wins across protocol quality and task interdependence, with the R* = κ

capacity threshold drawn in.

index.html

— the canonical browser simulation. It runs the scenario-driven Monte Carlo model where paper use cases map into the same abstract boundary levers: protocol quality, interdependence, observability, agent coverage, governance, market friction, volatility, and planning depth.

The earlier prototype pages now redirect to index.html

so old links do not open stale concepts.

paper/coase-information-theory.tex

— canonical working-paper sourcepaper/coase-information-theory.pdf

— rendered working-paper PDFpaper/coase-information-theory-draft.md

— earlier prose draft, retained for referencedocs/formal-paper-analysis.md

— original mathematical framing, propositions, agent extension, and completion plandocs/formal-note.md

— compact note tying the apps back to the theorydocs/paper-completion-plan.md

— concrete path from note to finished paperdocs/review-and-improvement-plan.md

— review of the current work and prioritized improvementsdocs/measurement-playbook.md

— how to estimate theory quantities from workflow tracesdocs/app-model-assumptions.md

— assumptions and calibration plan for the demo apps

@misc{guarraci2026boundaryagent,
  author = {Guarraci, Brian},
  title = {From the Boundary of the Firm to the Boundary of the Agent: Coase-Information Theory for AI-Mediated Organizations},
  year = {2026},
  howpublished = {Working paper},
  url = {https://github.com/galatheus-labs/coase-info-theory}
}

sample-data/incident-trace.csv

sample-data/support-ticket-trace.csv

These samples are retained for future calibration work.

The companion runs live at ** https://galatheus-labs.github.io/coase-info-theory/**. To run it locally, open

index.html

directly in a browser or serve the repo:

python3 -m http.server 8000

Then open http://localhost:8000/

.

The paper is strongest when it does five things together:

Formalizes agents as bounded information-processing units: individuals, teams, firms, vendors, and software agents.Explains boundaries as coalescing/splitting choices around surprise reduction and coordination cost.Defines agility as the rate at which information becomes coordinated action.Shows executable examples where software agents shift latency, distortion, monitoring, and the effective boundary of the firm.Runs a simulation where the formal objective produces boundary phase behavior across protocol quality and task interdependence.

The simulation is deliberately simple. It is not a calibrated causal estimate. It is an executable illustration and measurement scaffold.

Choose one empirical spine — preferably incident response or support routing — and replace the current stylized coefficients with real trace calibration.

This repo should support a public paper/essay package:

• arXiv-style working paper,
• companion demo repo,
• short founder essay,
• video walkthrough,
• one trace-based case study.

The thought-leadership sentence:

AI-native companies are not just companies with AI tools. They are organizations whose sensing, routing, decision, and execution loops are increasingly software-mediated.

source & further reading

github.com — original article