Three agents. One operating system.
An AI marketing operations system for SEO and Content — designed and run by the senior marketer who deployed it.
This is one of several stacks. The same architecture extends to paid acquisition, community, lifecycle, and analytics — built by an operator, not bought from a vendor.
Three agents, one pipeline. Index handles research at the input — keywords scored, clustered, approved. Press handles content in the middle — articles drafted, imaged, published. Wire handles distribution at the output — threads and posts produced, approved, scheduled. The connections matter as much as the stages: keywords move from research to content, articles move from content to distribution, nothing falls between the stages.
The agent that’s already running
Press — the content agent
The content agent is the only one of the three currently live. Built as a 21-node automation that takes a single keyword from a research queue and produces a publishable article — with original imagery — on the blog, end-to-end, without manual intervention between the brief and the live post.
The text pipeline has six stages: research input from a Google Sheet (keyword, content angle, target page), an architect node that drafts the article structure against editorial guidelines, a humanizer pass that rewrites roughly 40% of the output to break AI-pattern signals, a guard layer that checks for factual drift and tone violations, a code-based enforcer that does deterministic find-replace on remaining AI tells, and a publish node that pushes the finished article to WordPress as a scheduled post with focus keyword, meta description, and category assigned.
Imagery is generated inside the same pipeline, not added as an afterthought. Each article ships with a hero image; longer articles (1,001+ words) also get a body image at the 40% mark. Image prompts are templated against a fixed design narrative — the same dark-aesthetic, green-accent visual language used across the site — so every article looks like it came from the same publication, not a stock-photo grab bag. Each image carries SEO-clean alt text and a meta description, generated in a separate small node downstream of the image itself.
Model choices are operator decisions, not vendor defaults. Text reasoning runs on Anthropic — Claude Sonnet for the architect and humanizer (where rewrite quality matters more than speed), Claude Haiku for the guard and the alt-meta generation (yes/no classification and small JSON jobs where Sonnet would be overkill). Image generation runs on OpenAI’s gpt-image-1, which produced the brand-consistent output we wanted. The enforcer isn’t an LLM at all — it’s a Code node with a hand-curated find-replace dictionary, because some patterns (“delve,” “tapestry,” “in conclusion”) can be killed deterministically and shouldn’t burn tokens.
The agent is built around a non-negotiable operator constraint: every article carries a methodology disclosure linking back to this page. Readers know an AI agent drafted the piece. The disclosure is the editorial standard — if the output isn’t good enough to publicly attribute, it isn’t good enough to publish. That constraint is what makes the rest of the system work — the architect prompts, the humanizer aggression, the guard thresholds, the find-replace dictionary, even the image-prompt templates all get tuned against the disclosure standard, not against “does it sound human enough to get away with.”
PRESS doesn’t hand you a draft — it delivers. Read the output →
Research, without the third-party stack
Index — the research agent
The research agent is the input side of the system. Currently in design; the keyword queue that feeds Press is curated manually today. Index automates that step.
The brief is specific. Every two weeks, Index runs five jobs in sequence. First, it monitors a defined set of competitor domains and identifies the keywords they rank for that this site doesn’t — the opportunity surface. Second, it scores each candidate keyword on search volume, ranking difficulty, search intent, and alignment to the editorial angles this site actually publishes against. Third, it clusters related keywords so the Press pipeline can produce article series, not isolated one-offs — a stronger compounding pattern for topical authority than scattered single posts. Fourth, for the keywords that survive scoring, it drafts a Content Angle — the field the Press pipeline already consumes — navigated by a set of editorial guidelines the operator maintains. Fifth, the entire output lands on a dashboard for human review: one screen, every proposal, yes/no on each item, approved keywords flow into the Press queue automatically.
The constraint that defines Index is deliberate: no Ahrefs, no SEMrush, no Moz, no monthly subscription bleed for SEO infrastructure that costs more than the rest of the marketing stack combined. Most senior marketers default to those tools because they’re the industry pattern; Index replaces the workflow they automate, not the data they sell. The design favours the free, common, powerful infrastructure most marketing teams already have access to — Google Search Console API for ranking and impression data, Google Analytics for behavioural signal, public SERP analysis for competitive monitoring — combined with LLM reasoning over that data and the same model architecture Press uses. The unit economics shift: an SEO research engine that costs in pipeline tokens what a single Ahrefs seat costs in monthly subscription fees.
The dashboard is the human checkpoint. Index proposes; the operator approves. One screen, one click per item, no spreadsheet wrangling between research output and pipeline input. Same operator constraint as the rest of the system: the agent does the work that scales, the operator owns the editorial decisions that don’t.
Distribution, format-aware
Wire — the distribution agent
The distribution agent is the output side of the system. Currently in design.
Every time Press publishes an article, Wire picks it up and repurposes it for the two surfaces a senior marketing audience actually reads: X and LinkedIn. Same article, two formats, both tuned to the platform — not the same caption copy-pasted across both. Each piece runs through an architect-humanizer-guard pattern matched to the conventions of its platform: a thread on X that respects the platform’s rhythm, a long-form post on LinkedIn that earns the scroll.
The methodology disclosure carries through. Every thread and LinkedIn post Wire produces tells the reader the same thing the article does: AI drafted this post and the article it links to, and the operator who runs the account built the system that produced both.
The operator constraint applies here too — for now. Wire proposes the thread and the LinkedIn post; the operator reviews both on a dashboard, edits if needed, approves with one click, and Wire schedules the distribution. The agent handles the format work that scales without quality drift; the operator owns the editorial call on every post that ships under their name. The architecture supports full automation when the editorial guidelines mature enough to trust unattended — the human checkpoint is a design choice, not a technical limit.
The distribution gap is where most editorial investment dies: articles published, then under-shared, copy-pasted across surfaces, forgotten by the next week. Wire is the correction. Same operator constraints, same disclosure standard, same architecture pattern Press already runs — applied to the platforms that determine whether the work gets read.
What the agents are evidence of
Why these three agents are a showcase, not the destination
A senior marketing function has five operating workstreams: SEO and content, paid acquisition, community, lifecycle, analytics. Each one produces different work, on a different cadence, against different editorial standards. Each one is a candidate for the same operator-built treatment. Most of them won’t get it from a vendor — the editorial logic that makes the work credible doesn’t survive being abstracted into someone else’s product.
The three agents on this page operate one of those five workstreams. They run because SEO and content is the workstream with the most public-facing output — easiest to demonstrate on a portfolio page, easiest to verify by clicking through to the blog. The other four are not coming-soon features of this page. They’re the rest of the function, waiting for the same treatment when the role makes them the priority.
The page is a working sample, not a product catalog. What’s being demonstrated isn’t the agents. It’s the operating posture: that a senior marketing leader can hold the editorial standard, the production architecture, and the AI calibration in the same head — and ship working systems against that. The agents are how the posture becomes visible.
What this looks like inside a role
Looking for this posture on your team?
The three agents on this page run because building them was the cheapest way to demonstrate the work. The same posture, applied to the workstreams that matter to your function, runs faster — there’s a brief, a constraint set, and a team already in motion.
Get in touch →