“Agents are the orchestrators that define the processing power, and then skills and tools and memory are the resources that the agents use to perform that particular output.” — Lou
Session context: 2026-04-30_Mastermind — Lou presented the updated universal ambient intelligence architecture after studying Hermes and OpenClaw, framing it as a functional organizational model you build once and extend forever.
Core Idea
The most powerful thing you can do with AI right now is stop thinking about it as a tool and start thinking about it as an organization — one you design, one that runs on your computer, and one you own completely.
Lou’s architecture treats every folder as a functional unit. A “Company” folder holds your agents: Chief of Staff, Writing Team, Research Department, Copy Team. Each agent folder contains its own identity (SOUL.md), its own skills, its own memory, and its own tools. The Chief of Staff has a heartbeat — a loop that monitors whether anything needs to be done and orchestrates the right agents at the right time. The Writing Team has a Team Lead agent that dynamically calls Researcher, Writer, Editor, and Audience Avatar sub-agents in the right order.
The critical distinction: agents are orchestrators, skills are functions. An agent decides which skills to invoke and when. A skill performs the atomic work. A tool gives the agent access to the outside world — Bash, MCP, Python, computer use, browser access. Memory lives inside the agent’s folder, scoped to that function, loaded only when that agent runs.
This is a functional organization, not a hierarchical one. Each agent is expert in one function and can be composed with others for any task. The writing team doesn’t report to the research team — they collaborate when the Chief of Staff determines a task requires both. Add a new capability by creating a new folder with its own SOUL.md and skills. Remove one by deleting the folder. The intelligence travels with the file system.
What makes this design durable is what Lou called the adapter layer: each folder contains adapters for Claude, OpenAI, Gemini, and open-source LLMs. Nothing intelligent lives in the adapters — they only define how to talk to each provider. The intelligence lives in SOUL.md and the skills. If Claude becomes the wrong tool tomorrow, any LLM that reads files inherits your entire organization immediately.
The standard concern with autonomous agents — security risk, prompt injection, access to sensitive data — is handled architecturally: you control which tools each agent can invoke, and at what scope. An agent might get browser access limited to a single domain. Another might get Bash access but not network access. The organization runs autonomously within the bounds you define.
Practical Application
Build your first functional agent in one session:
- Create a folder called
my-company/ormy-team/. - Inside it, create one agent folder — start with
writing-team/. - Inside
writing-team/, createSOUL.md(who is this team? what is it for?), askills/folder with two skills (research.mdanddraft.md), amemory/folder (empty for now), and ateam-lead.mdagent that describes how it calls the skills in order. - Ask Claude to “run my writing team” on a topic you’d normally do manually. Watch how the team-lead agents calls each skill.
- After the first run, open
memory/and let the team-lead write a brief note about what worked. That’s your first piece of evolving organizational intelligence.
Coaching Question:
“If I were to design my business as folders instead of a job description — what would the top three functional units be, and what would each one’s one job be?”
Related Insights
- Insight - Ambient Intelligence — Build a Skill in Every Folder to Make Your Entire Knowledge Base Alive — the predecessor insight establishing the folder-as-intelligence principle; this insight extends it to full organizational design
- Insight - Separation of Concerns in Skills — One File, One Job — the same modularity principle applied at the skill level; agents apply it at the organizational level
- Insight - Skill Chaining — Build Modular AI Pipelines Instead of Monolithic Prompts — skill chaining within a pipeline; the functional org extends this into an orchestration model
- Insight - The Resolver Pattern — Your CLAUDE.md Is a Pointer File, Not a Knowledge Store — portability principle that the adapter layer in this architecture operationalizes
- Insight - Design AI Systems for Maximum Composability and Minimum Context Pollution — the design philosophy this architecture is built to enforce
- Insight - Process Architecture Transmits Judgment More Reliably Than Individual Prompts — why encoding your org in a folder structure beats ad-hoc prompting
- Insight - Decide the Abstraction Layer Before You Build — the abstraction-layer decision comes first; this architecture sits at the top of that ladder
Evolution Across Sessions
This builds on Insight - Ambient Intelligence — Build a Skill in Every Folder to Make Your Entire Knowledge Base Alive (2026-04-02), which established the baseline: a skill in every folder makes your knowledge base live. The new development is the organizational model — not just live folders, but a full company structure with agents as orchestrators, skills as functions, and adapters as LLM portability. The ambient intelligence insight was about knowledge; this insight is about operations. Future sessions should test which business processes get the highest ROI from this structure first.