My Agentic Loop

A company-like operating system for running a wide array of agentic LLMs, with me as the human in the loop.

AI Company OS — full architecture

32 AGENTS · 3 WINGS · 160+ SKILLS · 2 RAG STORES · 2 LIVE RUNTIMES · AS OF SEP 2026

Most AI setups are a collection of tools that happen to be in the same browser. I try to create a system and expand it: advisors that hold their domains and share what they decide, an executive layer that turns decisions into shipped work, an automation layer that moves work between them, and an evaluation layer that scores everything and proposes its own improvements.

It runs on cloud tiers and on some hardware I already owned for cost efficiency on certain tasks. It was built inside-out over seven layers.

This setup is continuously expanding. More to add in the upcoming weeks. Currently working on my personal and entrepreneurial GraphRAG.

Layer	What runs there	What it's for
Local	Llama 3.2 (3B), Qwen 3.5 (4B) and Qwen3 Embedding via Ollama — a spare laptop, 6GB GPU	A Free Play setup for various local LLM tests; LLMs can be replaced with others or can be extended with a different setup. Used mostly for Triage, classification, routing, reformatting, extraction. Simple and recurring tasks at zero marginal cost.
Free-tier cloud	NVIDIA NIM behind a gateway	Occasional use if needed as NIM is for development tests only. First-pass drafting, structured extraction, and the judge.
Premium	Claude, ChatGPT Codex, Gemini (Vertex AI), Qwen3.8 27B via OpenRouter	Core reasoning and the advisory layer. Claude, Codex and Qwen share the coding load. Gemini on Vertex AI leads marketing media. Load balanced.
Execution	Claude Code, with a full replica on ChatGPT Codex	The executive team — a head of orchestration, 3 wing heads and 29 specialists, a total of 32 agents sharing 160+ skills, all with scoped tool permissions and cost optimizations.
Knowledge	Two RAG stores on local embeddings	Grounded retrieval: an advisory archive and my personal knowledge base.
Evals	Arize Phoenix	Traces on every run, judge scoring and regression checks.
Automation	n8n	Watching folders, moving work between the two halves, scheduled runs.
Sync layer	Google Drive + Drive for Desktop	The shared state. Reachable by Claude, by Gemini, and as a plain local path by anything running on my machine.

The brain team. AI advisors with real functional identities — technology, product, marketing, finance, venture strategy. These are the ones I brainstorm with rather than delegate to. They hold context across sessions, they push back, and they coordinate through a shared folder rather than an integration.

AI Company OS — the brain team diagram

AI Company OS — the execution team diagram

The execution team. Where a decision becomes work. A head of orchestration sits on top of three wings: product, engineering, and marketing. The wings hold 32 agents, and those agents share more than 160 skills: research, specs, prototyping, code, tests, review, security, content, SEO, image and video generation, and documentation work. Safety here is enforced through tool permissions. Agents are mixed of models and providers. Orchestration is mainly led by Claude Code, but a full replica is available to run through ChatGPT Codex, so work can move from one to another when session limits run low or when a task needs a second engine.
The knowledge layer. Two RAG stores sit beside the shared memory: an advisory archive and my personal knowledge base. Both run on a local embedding model, so each lookup costs nothing. Agents retrieve grounded context from them instead of guessing.

The nervous system. The automation layer plus a shared file-based memory that both halves read and write. Work crosses between thinking and doing without a human carrying it.

AI Company OS — the nervous system diagram

AI Company OS — evals and observability diagram

Evals and observability. Arize Phoenix traces every run, and a judge model scores output against rubrics calibrated against real outputs. Regression checks sit here too, so improving one agent doesn't quietly break another.

The self-improvement loop. Eval findings become proposals — a prompt edit, a rubric change, a model-tier promotion — queued for approval. Approved changes are applied, and that agent is permanently better. Fifteen to thirty minutes of review in a morning.

AI Company OS — the self-improvement loop diagram

AI Company OS — external reach diagram

External reach. Analytics, ad platforms, video and email APIs. Last layer in the whole system. Expands or shrinks from time to time due to active needs.