Skip to main content
SEO here is operated through an AI agent (Pi) rather than by a human clicking through tools. The operator’s job moves from doing the steps to directing an agent that does them in parallel with subagents. This is the “Next” era from 1.2 turned inward: the human gives a task, and the agent fans out the queries and the work. The leverage is the whole 4-week loop running as delegated agent work with durable memory.

​
Pi is the surface you operate

“The AI agent that’s central to entire operations… what you’ll mainly deal with 90% of the time.”
Work is issued as a task, and Pi decomposes it (“do in sequence, passing data from prev to the next, use subagents”), so research, product bible, keywords, and pillar planning run as a chained agent pipeline instead of a human clicking through screens. State persists in a virtual filesystem (/memory/[company]/research.md, product-bible.md), so the agent’s knowledge compounds across the loop rather than living in an operator’s head. The same agent’s-eye view that builds is also used to audit the market: Birdseye shows how Claude Code searches, which closes the loop between “how agents search” (1.2) and “how we build for agents.”

​
The canonical operator task

This one message drives a whole onboarding; it’s the prompt we run live: 
I'm working on pixo.video, help me:
1. do a website research
2. build a product bible
3. update my org info
4. do deeper research on what bofu keywords I need, see 2d keyword labeling framework
5. expand on the keyword and think about programmatic seo templates
6. (in parallel to 5) using the product bible help me plan a horizontal landing page pillar (create it).
do in sequence, passing data from prev to the next, use subagents.
Pi runs the steps in order, hands each step’s output to the next, and parallelizes where the tasks are independent (5 and 6). Other real one-line operator commands work the same way:
  • “Do a website research” · “Build a product bible”
  • “look at ALL the keywords right now and run it through the keyword evals”
  • “load all keywords into synscribe… tag all the clean accept keywords as ‘Core’”
  • “Do a technical SEO audit on [site]” · “setup my image generators / setup my default refiners.”
The work products land in Pi’s working directory (/memory/[company]/research.md, product-bible.md). You read the agent’s output rather than producing it by hand.

​
Where the leverage actually is

The agent does more than 90% of the manual execution, from research to shipping code changes to a client’s live site, so a human’s job collapses to quality gates and the client relationship rather than production. That’s what lets one operator carry many clients, and it compounds along four lines:
  • Parallelism. Subagents run independent steps at the same time, and self-learning agents run parallel experiments across the whole org.
  • Consistency. SOPs become upgradable skills the agent runs identically every time, like the automated indexer that submits roughly 150 pages per client per month with no one lifting a finger.
  • Compounding memory. Each client’s /memory/[company]/research.md and product bible carry everything the agent knows about them forward, so context doesn’t decay between tasks.
  • Compounding capability. Every base-model or harness upgrade raises the ceiling for the whole fleet at once.
❓ [needs Raymond: the concrete leverage ratio]

​
Operationalize it