The Compounding Organization

Abstract

NOAERTH began with the act of building. Projects were pursued not only as potential businesses or products but as attempts to answer questions. Each attempt exposed constraints and possibilities that could change what we tried next. As the number of projects grew, the founder became the limiting resource. AutoBuilder emerged as an effort to delegate smaller tasks to AI and preserve the founder's attention for strategy, judgment, and new ideas.

This white paper develops that experience into a broader proposition: an organization can potentially expand its future capacity when its projects produce not only outputs, but verified and reusable knowledge. We call this a compounding organization. Its central process is a repeating, expanding sequence: build, learn, understand, question, and build again. We describe a research architecture, the portfolio as a collection of learning environments, the role of failure, the evolution of AI delegation, and a plan to test whether reuse actually improves new work.

1. Building Before the System

The first principle is simple: making something changes what we know about it. Ideas become more concrete when they encounter code, customers, production constraints, artistic choices, or real-world feedback. The builder learns what was missing in the original question.

From the beginning, NOAERTH has treated building as a way to discover. That generated more ideas than one person could sustain. Instead of concentrating exclusively on making one startup succeed, the founder experimented across many possibilities, accepting that some would fail while others might reveal a better direction.

That approach presented a practical limit. Every new project brought routine work: implementation, debugging, organizing files, writing tests, setting up infrastructure, checking results, and revisiting unfinished tasks. The person choosing what to build was also doing too much of the building.

AutoBuilder was conceived to address that bottleneck. Its ambition was an always-available builder that could make progress on small, repeatable tasks while the human focused on the larger view. At the time, AI's useful range was narrower; the need for human checking was obvious. As AI capabilities developed, so did the possible scope of delegation—from small changes toward larger tasks involving planning, coding, coordination, and review.

The wider question soon became more important than any individual product: what would it take for an organization to keep building, keep learning, and carry forward what it understands?

2. What Compounds?

The growing number of projects is not itself the measure of progress. Ten incomplete products may teach very little. One difficult project might produce a method or component that benefits many later attempts.

We distinguish three kinds of output that can emerge from a build: an artifact (something that exists), an insight (something understood), and a new question (something worth investigating next). An organization compounds only when useful parts of this output are preserved and applied appropriately.

To be useful, organizational memory must be selective and corrigible. A lesson learned from one codebase may be wrong for another; a design that worked with an older model may become obsolete. The system must allow people to reject, revise, or forget misleading conclusions.

3. Projects as Learning Environments

NOAERTH's portfolio spans autonomous agents, software-building systems, operating tools, games, interactive worlds, and music production. These are not all equally mature, nor are they all evidence for the same claim. Their value to the research vision is that each exposes different constraints.

AgentOS and related systems ask how agents should coordinate and how completion should be verified. PAIOS and storage work ask what an organization should remember. REALM and FRIENDS expose the difficulty of building complex, stateful interactive environments. AudioOS and MusicOS raise questions about evaluating and revising creative outcomes. CompanyOS and AutoERP test whether methods designed for technical work can inform operations.

The projects become a research map when we document what prompted each one, what was attempted, what it revealed, and which later decision it influenced. That record must distinguish remembered inspiration from relationships supported by dated code, notes, and releases.

4. Failure as Information

Rapid experimentation is sometimes described as failing fast. In NOAERTH, the intended meaning is more specific: make an attempt early enough to discover whether its assumptions hold, then use the result to change what happens next.

A venture can fail commercially while revealing a strong technical pattern. A simulation can look unconvincing and expose an important production constraint. A software agent can produce plausible output that fails a real test. These outcomes are only useful if the reasons are investigated.

A compounding organization should therefore treat failed attempts as first-class records, without romanticizing them. It needs to know when to reuse a lesson, when to reject it, and when it has insufficient evidence to generalize.

5. AI Changes the Shape of the Organization

AI did not create the desire to build. It changed the possibility of how much work could be delegated. Early models were most useful on smaller tasks and needed close supervision. More capable models and agent systems now make larger, connected tasks feasible to attempt, but do not remove the need for review or clarify responsibility by themselves.

We propose a human-directed operating model with four functions: set direction, organize and execute work, verify outcomes, and preserve useful knowledge. The distinction between an agent's report and an independently evaluated result is essential. Otherwise the system can accumulate confidence faster than competence.

The long-term possibility is an organization that can conduct many experiments without forcing one person to personally perform every repetitive step. This is not a proposal to remove people from judgment, creativity, or accountability; it is a proposal to give human direction greater reach.

6. A Future of Smaller, More Capable Organizations

Imagine someone arriving with an idea that would once have required a large team just to investigate. A shared system could search earlier work, assemble tools, suggest a plan, build a first version, test its assumptions, and record the result. The next project could start with those verified assets rather than from a blank page.

That future matters beyond software startups. Similar principles could support small organizations, scientific experimentation, educational tools, interactive worlds, and public-interest applications. But greater output is not the same thing as greater benefit: costs, access, accountability, security, and who bears the risks all matter.

The ambition is a different relationship between people and technology. Human curiosity initiates the work. Tools extend what can be tried. Experience improves understanding. Understanding shapes the next act of creation.

7. How We Could Test the Thesis

The central claim is deliberately testable: verified knowledge from earlier work may improve performance on later tasks that were not known when the earlier work was recorded. A first study could compare task-matched agent workflows with and without curated cross-project memory.

We would examine independently verified completion, development time, model and tool cost, human intervention, recovery from errors, and whether retrieved knowledge helped or harmed the new task. Trials should include unsuccessful results and require comparable budgets and clear stopping rules. A related study could isolate the effect of multiple agents from the effect of model routing.

We do not yet treat this as a demonstrated compounding effect. If stored knowledge produces irrelevant suggestions, greater complexity, or additional review burden, the outcome would revise the idea. The NOAERTH research agenda describes the proposed comparisons.

8. Intellectual Foundations

Learning through productive activity has a long intellectual history. Kenneth Arrow studied the economic implications of learning by doing [1]. David Kolb described learning as a cycle connecting experience, reflection, understanding, and new experimentation [2]. James March examined the balance between exploring new possibilities and exploiting what is already understood [3]. Ikujiro Nonaka explored the creation and transfer of organizational knowledge [4].

More recent work on coordinated language-model agents and software evaluation informs the technical parts of the NOAERTH proposal [5–6]. The proposed contribution is not to claim these ideas are new in isolation, but to explore their integration in an evolving human-directed building environment.

References

1. [1] Arrow, K. J. (1962). The Economic Implications of Learning by Doing. The Review of Economic Studies, 29(3), 155–173.
2. [2] Kolb, D. A. (1984). Experiential Learning: Experience as the Source of Learning and Development. Prentice Hall. University of Florida overview.
3. [3] March, J. G. (1991). Exploration and Exploitation in Organizational Learning. Organization Science, 2(1), 71–87.
4. [4] Nonaka, I. (1994). A Dynamic Theory of Organizational Knowledge Creation. Organization Science, 5(1), 14–37.
5. [5] Wu, Q., et al. (2023/2024). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation.
6. [6] Jimenez, C. E., et al. (2024). SWE-bench: Can Language Models Resolve Real-World GitHub Issues?.