SEOs were running agent harnesses a decade before the word existed. The executors were people, the prompts were briefs, and the eval suite was an editor with a checklist.
Author: James Phoenix | Date: July 2026
I Bought Articles for $20. Links Went for $40.
When I worked in SEO, scaling meant buying execution. Articles cost me between $20 and $80 depending on the niche and how much the writer pretended to know about it. Around me, people bought placed links at roughly $40 each. A modest campaign of thirty articles and twenty links was a $2,000 to $3,000 monthly line item, and that was the cheap version, because the alternative was doing it all yourself.
The workflow was always the same. I wrote a brief: target keyword, word count, headings, internal links, tone, things to avoid. I sent it to a marketplace or a freelancer, usually someone several time zones away working for a fraction of a UK salary. Drafts came back. Some were usable, some needed an editor to rebuild them, some went straight in the bin. A spreadsheet tracked every URL, status, and payment. When something weird came back, it escalated to me.
Nobody called this an agent system. It was an agent system. We just hadn’t invented the vocabulary yet.
The Machinery Was Already a Harness
Map the 2015 SEO operation onto 2026 agent infrastructure and every component has a counterpart:
The humans were the inference layer. Everything around them, the part that actually made the operation work, was orchestration: decomposing the goal into brief-shaped units, constraining the output space up front, checking results against acceptance criteria, retrying failures, and routing exceptions to someone with judgement.
That orchestration layer was the hard part and the expensive part. The writing itself was, by design, the most commoditised component in the system. We had already engineered the work so that the executor was interchangeable.
Why SEO Flipped First
When GPT-3.5 landed, the SEO industry converted almost overnight. Not because SEOs were unusually visionary, but because the migration was trivial. The work was already text-in, text-out. The outputs were already checkable against a brief. The quality bar already tolerated probabilistic output, because a $20 article from a stranger was never a sure thing either. Swapping a human executor for a model executor changed one component and kept the whole harness intact.
Intelligence was never the bottleneck. Decomposition was. The decade of outsourcing forced SEOs to do the genuinely hard work of automation before the automation existed: break fuzzy goals into units a stranger can execute from a written spec, define what “done” looks like, and build the QA and escalation loops that catch failures. GPT-3.5 didn’t automate SEO content. SEOs had already automated it. The model just undercut the previous executor on price and latency.
This explains something that confuses people about agent adoption today. Industries insist they are “waiting for the models to get good enough,” but the models have been good enough for most of their tasks for a while. What those industries are actually missing is the decomposition. Their work has never been broken into brief-shaped units with checkable outputs, so there is nothing to hand an agent. They aren’t behind on AI. They’re behind on SOPs.
Offshore SOPs Are the Map of What Agents Eat Next
This gives you a falsifiable way to predict agent adoption order. Ignore the demos. Look at what currently runs on offshore SOP labour: lead research and enrichment, data entry between systems that lack APIs, content moderation, listing management, bookkeeping preparation, tier-one support, template-driven video editing, outreach coordination.
Every one of those already passed the decomposition test. Someone proved the task can be executed by a stranger with a checklist, at arm’s length, with output QA’d against acceptance criteria. That is precisely the shape of work an agent slots into with no re-engineering.
The prediction: agent adoption tracks decomposition, not model capability. If a process can be handed to an offshore team, it can be handed to an agent, and it will be, in roughly the order the outsourcing industry already sorted for us. Work that resists outsourcing, the ambiguous, political, trust-heavy, responsibility-bearing work, resists agents for exactly the same reasons, and no model release changes that until someone does the decomposition.
The Arbitrage Dies Twice
Here’s the part the “agents are the new offshore labour” crowd leaves out, and it’s the part I watched happen.
Cheap human execution didn’t just scale SEO content. It flooded the channel until Google repriced it. Panda in 2011 wiped out the content farms that $8 articles built. The playbook survived, moved upmarket, got more careful, and then AI made execution nearly free, so everyone ran it again at a thousand times the volume. Google’s 2024 spam and helpful content updates repriced the channel a second time, faster and harder.
Same loop, same physics: when execution gets cheap, the channel it exploits degrades until the arbitrage is dead. Swapping executors doesn’t save an arbitrage. It accelerates its funeral, because the flood arrives sooner. The outreach world is running this experiment right now, with agents drowning inboxes the way content farms once drowned search results, and it will end the same way.
So the honest lesson from SEO isn’t “now we can scale without the humans.” It’s twofold. The harness you build, the decomposition, the evals, the escalation paths, is durable and transfers across executors: humans, GPT-3.5, whatever comes next. The arbitrage it executes is not. Content survived as a task and died as an arbitrage, twice.
Which is why I don’t buy thirty briefs a month anymore, from humans or from models. When execution costs nothing, everyone has execution, and the only scarce input left is having something to say.
