I went from outsourcing SEO production to running an autonomous agent workforce.
In 2023 I bought SEO by the task. By 2026 I design, build, and operate a fleet of AI agents that ship real work on production systems — technical SEO, content, and business infrastructure — supervised by one operator through hard guardrails. This is the documented record.
One operator. A documented system.
The operating model behind everything on this page — verified against the live Agentic Brain vault and the project record.
One human sets direction. An orchestrator routes work to a self-hosted fleet of agents, each with a defined role. Every agent reads and writes a governed, git-backed shared memory before acting, drives a real toolchain, and ships work to production on the Tier 1 path. Anything ambiguous stops and escalates to the operator on the amber Tier 2 loop. Backups before edits. Write, then verify. Every action on the record.
Every system on this page links to a full technical dossier — the problem, the architecture, and what actually shipped.
Orchestrator, workers, and local models running on machines I own — coordinated in Slack, supervised by one operator.
A git-backed vault every agent reads before acting — with write-governance enforced by rules, not vibes.
Safe work ships on its own. Anything ambiguous blocks until a human approves it. Backups before edits, write-then-verify.
The flagship SEO engagement (May 2022 – April 2026).
Shown to frame the surface the work operated on. These are engagement-wide outcomes across a four-year client relationship — not attributable to any single agent, sprint, or automation.
Engagement-wide context only. These outcomes reflect a four-year client relationship and are not claimed as the result of a single system on this site.
What running the machine takes.
The capability stack behind the portfolio, grouped the way a hiring team reads it — from agent design through to full-stack builds.
AI Agent Design & Orchestration
Automation & Reliability
Memory & Knowledge Systems
Technical SEO & AEO / GEO
Full-Stack Builds & Infrastructure
Operating Discipline
The documented work.
Three tiers, honestly sorted — the flagships that carry the story, the systems underneath them, and the smaller builds and experiments kept for breadth. Every card links to its full technical dossier.
The systems that carry the story.
The flagship builds — deepest, most-documented, and closest to the client results above. Each links to its full technical dossier.
Keystone — an autonomous technical-SEO agent running an enterprise site
An AI agent that ran production technical SEO on an enterprise WordPress VIP site — shipping real code, repairing broken link networks, deploying and validating schema, and root-causing a hidden cache regression, on a daily cron under a two-tier human-approval model.
The Agentic Brain — a second brain my agents read
A git-backed Markdown vault that is the shared long-term memory for me and my agents — Obsidian for humans, GitHub as the source of truth, four contribution states, and an auto-compiling knowledge wiki so a writable brain can't rot.
The Content Machine — library-scale publishing, compliance-governed
An AI-assisted publishing operation that turned a high-authority health brand's own video library into topical-authority articles at scale — every piece run through editorial and compliance review, with a compliance-driven pivot that kept the record clean.
hermes-jobsearch — an agentic job-discovery pipeline
A scheduled pipeline that pulls SEO/GEO/AEO roles from the major job boards, dedupes and hard-filters them, verifies every apply link, scores each survivor with a local model, and emails a ranked digest — later rebuilt into a multi-phase intelligence system.
A multi-agent fleet across three Mac minis
A team of AI agents — an orchestrator, an ops agent, a sovereign worker, and a local LLM — that coordinated in Slack, handed off Google Workspace credentials, and provisioned SSH access between machines, running a real job-lead pipeline across the fleet.
The infrastructure underneath.
Orchestration, memory, model routing, self-hosted compute, AI-search plumbing, and the business builds — the load-bearing layer the flagships run on.
Master Control
An always-on OpenClaw orchestrator that owned a cron fleet, coordinated sub-agents and workers, reported daily, and escalated anything risky for approval.
View →Built to Be Cited — AEO / GEO
The AI-search infrastructure that makes a high-authority health brand citable by ChatGPT, Perplexity, Gemini, AI Overviews, and Copilot — from llms.txt to MedicalWebPage schema.
View →Model Routing & LLM Migration
Different LLMs for different jobs — frontier models for reasoning, cheap models for crons, local models for volume — plus a cost-driven migration off the Anthropic API.
View →Three-Mac-Mini Cluster
A self-hosted cluster of three Mac minis running local LLMs and an agent fleet — networked over Tailscale and Thunderbolt, moved cross-country, and honestly audited to a hardware wall.
View →The Memory Stack
A layered memory system so self-hosted agents stop forgetting — lossless context, structured long-term recall, a company knowledge brain, and a git-backed vault across three machines.
View →The LLM Wiki
Karpathy's LLM Wiki pattern inside my vault — immutable raw-source intake, a hashed source registry, and an AI-compiled, lint-checked knowledge graph.
View →Slack Agent Coordination
Wired MC, Hermes, and Claude to coordinate as distinct identities in Slack — then diagnosed and patched a real threading bug over SSH, with a test and a commit.
View →Paperclip — Control Plane
Agents run on a real control plane, not vibes: every task routed by identity, every "done" gated by QC, every action written to an audit trail, issues tracked as HOL-### tickets.
View →The Cron Layer
A self-maintaining automation layer — scheduled watchdogs, keepalives, git-backed memory sync, and briefing jobs — audited hard enough to prove a local LLM couldn't hold a production cron.
View →holisticseo.co Rebuild
Killed my old WordPress site and rebuilt holisticseo.co from scratch as a Next.js app on Vercel with a Sanity blog and Cloudflare DNS — shipping in days, solo, with Claude Code.
View →agenticworkforcecompany.com
Built from scratch on Next.js / Vercel / Cloudflare — a portfolio-and-CV site that names the exact agent infrastructure I run and positions the AI-era operator story in one place.
View →Job-Application Engine
A governed answer bank of every application question with approved answers, reusable cover-letter hooks, and an agent that drafts and posts my LinkedIn work history.
View →Kept for breadth — and honesty.
Early, one-off, and experimental builds. Some shipped and did their job; two I deliberately shut down when they stopped earning their keep. Knowing what to abandon is an operating skill, so they stay on the record.
Not flagships — but each is real, documented, and honestly labeled. The amber tags mark builds I chose to retire.
- The Briefing Layer →Scheduled agent jobs — a morning ops briefing, an overnight X intel digest, a daily debrief, and a live master to-do feed.Iterated
- Overnight Transcription →Pointed an agent at a paywalled, caption-less course — it pivoted to local Whisper and ground through the entire course overnight, delivering word-for-word transcripts of every lesson.Shipped
- On-Demand Documents →One chat message → a branded Certificate of Employment: letterhead built, the three questions that mattered asked, editable Google Docs plus sign-ready PDFs delivered.Shipped
- The Ops Dashboard →Built a dashboard to watch my agents and crons — then, when it stopped earning its keep, deliberately let it go instead of babysitting it.Abandoned
- Trading-Signal Experiment →A multi-agent market-signal pipeline — multiple data feeds, a convergence scorer, autonomous paper-trading — measured honestly, found to have a losing edge, and shut down. Zero real dollars risked.Abandoned