The Org Chart You Have Never Seen Before
Traditional companies organize around humans with titles. Our company organizes around agents with instruction files. Every agent has a CLAUDE.md — a detailed document that specifies its role, its inputs, its outputs, its success criteria, and its constraints.
These instruction files are not suggestions. They are the agent's operating system. When the CEO agent starts a cycle, it reads its own CLAUDE.md line by line and executes each step.
The instruction files are also mutable — and this is where it gets interesting. The Auditor agent can rewrite any other agent's instructions if that agent is consistently underperforming. The system literally rewires itself.
The CEO: Strategic Orchestration
The CEO does not write code, design brands, or produce content. It decides. Every cycle starts with the CEO loading context: the registry, the git log, the analytics report, and the auditor's last report.
Based on this data, the CEO makes three types of decisions:
- Build — commission a new site from a Research-approved niche
- Improve — allocate agents to fix or enhance an existing site
- Deprecate — shut down a site that is not performing
The CEO also enforces the ratchet: if the portfolio score dropped since last cycle, no new builds. The system must fix what is broken before expanding.
The Auditor: The Immune System
The Auditor is arguably the most important agent in the system, and it is deliberately the last to run in every cycle. Its job is quality control at the system level.
After all other agents have completed their work, the Auditor reviews everything:
- Did the Builder's sites actually deploy? Are they returning 200 status codes?
- Did the Content agent's articles meet quality standards?
- Did the SEO agent configure all GEO endpoints correctly?
- Did the Research agent's niche scores align with actual performance?
Each agent receives a grade: A (excellent), B (good), C (needs improvement), D (failing). For agents graded C or D:
- First occurrence: the Auditor proposes instruction changes in its report
- 3+ consecutive weeks of the same failure: the Auditor directly edits the agent's
CLAUDE.md - If the instruction change improves metrics next cycle, it stays. If not, the Auditor reverts it.
The critical design choice: the Auditor reviews itself too. If its own recommendations led to worse outcomes, it adjusts its own evaluation criteria. This recursive self-improvement is what makes the system anti-fragile.
The Specialists
The Research agent does not just find opportunities — it tracks how accurate its predictions were. If it scored a niche at 85 but the resulting site underperformed, it adjusts its scoring model. It also maintains a knowledge/rejected/ directory of niches it investigated and rejected, with reasons — preventing the system from re-researching the same dead ends.
Every site in our portfolio has a unique visual identity. CalcFit uses clinical blues. PayScale Pro uses trust-signaling navy. Pixelry uses creative magentas. The Designer produces a complete brand JSON spec that the Builder consumes — ensuring the design intent is perfectly preserved in code.
The Builder is our most prolific agent. It takes a design spec and produces a complete, production-ready Next.js site. Every calculator implements validated formulas. Every tool processes data correctly. The Builder verifies its own work with curl checks after deployment.
The Virtual Newsroom
This is perhaps the most unusual part of our operation. We run a virtual newsroom with five named writer personas, each with a distinct editorial voice and subject matter expertise.
Why personas instead of a single generic voice? Because tone matters. A fitness article needs a different voice than a finance article. A trend explainer needs different energy than a technical analysis.
These are not fake people. We are not pretending they are human. They are personas — consistent editorial voices that produce better content than a single generic voice could. Each persona's byline appears on articles they “write,” and readers can expect a consistent tone from each one.
The Operations Backbone
Analytics runs first in every cycle for a reason: every other agent's decisions depend on having current data. The Analytics agent computes a health score for every live site — a composite metric based on uptime, page performance, traffic trends, and engagement signals.
The SEO/GEO agent has a unique dual mandate. Traditional SEO is table stakes. But we also optimize for GEO — Generative Engine Optimization. Every site exposes /llms.txt endpoints that AI assistants can read.
How They Work Together
The agents do not work in isolation. They read each other's outputs. The CEO reads the Analytics report and the Auditor's recommendations. The Builder reads the Designer's brand spec. The Content agent reads the SEO agent's keyword recommendations.
This is coordination through shared state rather than real-time communication. The registry (registry.json) and the git history serve as the shared memory layer. It is a fundamentally different model from how human teams work. No meetings, no Slack channels, no status updates. Just structured data flowing between specialized functions.