Four AI Agent System that turned 40 hours of specialist work into 4 hours of operator time.

Problem

The client is a digital marketing agency serving medical aesthetics clients. Their core service, producing SEO-driven website content, had become a bottleneck. A single page required eight roles and three to four weeks to produce. Writing alone consumed the first week.

I was brought in as a Product Designer. After delivering a design system and two website redesigns, I was given a broader mandate: diagnose and fix the content production process.

When I mapped the existing workflow, I found the real problem: every person in the pipeline was working from a different version of the client. No shared source of truth existed. None of the AI tools the team had experimented with had consistent access to reliable data, and in a medical aesthetics context, where compliance language and clinical claims carry legal and reputational risk, that made generic AI output unsafe to publish. The prompts were not the problem. The information underneath them was.

This was an information architecture problem. The process had not been designed. It had accumulated.

Approach

The four-agent system did not arrive fully formed. Each agent was added in direct response to organizational changes.

1. The Writer Agent came first. It targeted the most expensive bottleneck, the writing phase, and proved that a structured Client Bible plus a defined brief produced publication-quality drafts.
2. The Proofreader Agent came next to reduce the cost of human editing rounds. A scoring rubric had been on the team’s wishlist for some time but had never been built. The Proofreader was where I implemented it: enforcing the rubric so quality was measurable, not just felt.
3. The Strategy Agent came third when the strategist left the agency. This was the moment the agents stopped being a toolkit and became a pipeline.
4. The SEO Agent came last as a natural extension once the strategic and editorial layers were stable.

One agent, one job. Strategy decides what to write. Writer drafts. Proofreader diagnoses against a rubric. SEO optimizes. Strict separation made the system testable: when something went wrong, I could trace the failure to a single agent and a single input.

Hierarchy, not flatness. The Strategy Agent is the spine. Every downstream agent depends on the Strategic Brief it produces. If the brief is wrong, everything downstream is wrong. That is a feature, not a flaw.

Humans at every gate. No agent hands off to another agent. Every output passes through a human reviewer before the next agent runs. In a medical aesthetics context, where compliance language and clinical claims carry real risk, this is non-negotiable.

Pivot

The pivot came late and under pressure.

The entire system had been designed and built on ChatGPT GPTs with the Client Bible structured in Google Drive. Mid-project, the company standardized on Claude and Notion across internal tooling. The platform I had built on would not integrate. A budget was approved and I rebuilt.

The pivot was not a translation. It was a rebuild.

The conceptual model held. The source-of-truth principle, the ownership separation, the mandatory sequence, and the human gates all stayed intact. What I had built was a system, not a tool stack. The platform was a layer underneath it, and the platform changed.

Everything below the conceptual model was rebuilt from scratch: the knowledge layer, the agent configurations, the system prompts, the prompt libraries, and three MCP integrations with their own authentication and credentials handling.

I had not built MCP integrations before. I learned, configured, and shipped. I also wrote the operator-facing setup documentation for a non-technical marketing team, covering installation, credentials configuration, and known failure modes.

System

The system has two layers: a structured knowledge base called the Client Bible, and four AI agents that consume it in a fixed sequence.

The Client Bible is the canonical source of truth for every client. It lives in Notion as a structured collection of topic-specific files, each with a defined purpose, a defined owner, and a defined audience among the four agents. I designed the structure, the file requirements, the standards, and the governance rules. The Accounts team became the custodians. I owned what made their work usable downstream.

Each agent is a Claude Project with a defined system prompt, a dedicated prompt library, and configured MCP integrations. Each one reads only the files it needs and produces only the outputs it is responsible for.

Strategy Agent. Outputs the Strategic Brief: page purpose, outline, keyword strategy, internal linking, and notes for every downstream agent.

Writer Agent. Receives the Strategic Brief. Outputs a first full draft balancing brand voice, compliance, technical accuracy, and SEO requirements.

Proofreader Agent. Receives the draft and the brief. Outputs annotated corrections, a compliance and tone check, a scoring rubric, and a Ready-for-SEO decision.

SEO Agent. Receives the cleaned draft. Outputs metadata, URL recommendations, internal linking, schema suggestions, and cannibalization warnings.

The order cannot be skipped, and every handoff passes through a human reviewer before the next agent runs.

During the pilot, I noticed the system was slow. Claude was spending too much time searching for files and prompts at runtime. I diagnosed it as a retrieval pattern issue and built two indexed lookup tables: one mapping client files by ID, one mapping prompts by location. Both replaced open-ended search with direct lookup. The fix is documented as an SOP the team can maintain.

Selected documentation is linked below. The full library was delivered as part of the project handoff.

• How the Agents Use the Client Bible — the agent-to-file matrix that defines what each agent reads. The strongest single visual artifact of the architecture.
• SOP: Using the Client Bible ID Registry for Agent Setup — the registry SOP, evidence of the iteration loop that responded to the pilot.
• How to Set Up Claude Desktop with MCP — operator-facing technical setup documentation, written for a non-technical team.

One request moved through every stage of the pipeline during the pilot with the agency’s most complex client. Each artifact below is an actual output from that run.

1. Strategic Brief — the Strategy Agent’s output. 
2. Type 2 Heavy Refresh Draft — the Writer Agent’s first full draft.
3. Proofreading Report — the Proofreader Agent’s annotated diagnosis.
4. Compliance Revision Draft — the corrected draft after the human review gate.
5. Compliance Verification Report — the Proofreader’s second pass confirming corrections held.
6. SEO Optimization Report — the SEO Agent’s output.

Six artifacts. Four agents. Three human gates. One request.

Outcome

The system delivered a 90% reduction in content production time, from one week to four hours per piece, from request to client-ready draft. Four specialist roles were absorbed by four agents. One operator now runs the pipeline.

The pilot ran against the agency’s most complex client. The architecture held under real conditions. Content was produced, proofread, optimized, and delivered to the client for approval on schedule.

The complete Notion library, every SOP, template, and standard needed to build a Client Bible and configure the four agents from scratch, was delivered as part of the handoff. A new operator came up to speed using only that documentation, without me in the room. That was the test it was designed to pass.

Reflection

The platform pivot was expensive for the company: budget spent twice, and pilot runs I did not get to complete before I left. Early in the project I raised concerns about the existing tooling and proposed evaluating alternatives. The decision was to proceed. Mid-build, once the system was visible to adjacent teams, the budget opened up.

What I would carry forward is pushing earlier for the people who hold the budget to see the work alongside the people who will operate it. Raising it in the right room at the right time is part of the design work for a system this cross-functional.

AI systems are information architecture problems first and prompt engineering problems second. The earlier attempts at this agency failed because they treated AI as a content generator and kept trying to fix it with better prompts. The prompts kept failing because the information underneath them was fragmented.