I Built My Own AI Content Team: Content Manager, CMO, and CTO Agents

The first thing you learn when you start running a portfolio site as an actual business is that content is the job that never finishes.

Blog posts need drafting. LinkedIn posts need writing. The pipeline needs checking. Keywords need tracking. And all of this competes with the actual development work that pays the bills.

My solution was not to hire. It was to build.

The Three Agents

The agent system lives inside the santoshpaudel.me admin panel under the Agents section. Three agents run on a weekday schedule, each with a defined role:

Content Manager — Reads the current blog post queue, identifies topic gaps based on the keyword strategy, and proposes new posts. Each proposal includes a title, target keyword, excerpt, and a full draft. The draft goes into the pending queue — I approve or reject before anything publishes.

CMO (Chief Marketing Officer) — Reviews recent client interactions and newsletter metrics, identifies follow-up opportunities, and proposes tasks. If a prospect has not heard from me in 14 days, the CMO proposes a check-in. If newsletter open rates dip, it proposes a subject line test.

CTO (Chief Technology Officer) — Reviews the codebase and pending GitHub issues, proposes improvements, flags technical debt. Less active than the other two, but useful for catching things I would otherwise defer.

The Architecture

Each agent is a row in the agents table with a slug, a name, a system prompt, a provider, and a model. The agent runner pulls the agent config, gathers context from Supabase (recent posts, client pipeline state, newsletter metrics, the knowledge base), calls the LLM API with tool use, and logs the results.

The critical design decision: agents never write directly to business tables. Every output goes to pending_actions with status pending. I review at /admin/agents/pending and either approve (which executes the real insert) or reject. The agents propose. I decide.

The Models

I mix models based on task type, not loyalty:

Llama-3.3-70B via Groq — Powers the Content Manager for draft generation. Groq's inference is fast enough that a 1,500-word blog draft takes about 8 seconds. The output quality on long-form content is strong with a well-crafted system prompt.

DeepSeek R1 — Powers the CMO for reasoning tasks — identifying patterns in client data, generating follow-up logic. DeepSeek R1 handles multi-step reasoning well and is significantly cheaper than comparable models.

Anthropic (Claude) — Used for tool-use runs that require reliable function calling. Tool use on other models is inconsistent. For the agent runner's tool-use loop, this is not the place to optimize for cost.

The Actual Cost

Total AI spend across all three agents running daily on weekdays: approximately $0.0035 per month.

That number sounds wrong. It is not. The agents run on short context windows with efficient prompts. The Groq tier for Llama is extremely cheap at scale but trivially cheap at this volume. DeepSeek pricing is a fraction of GPT-4 pricing for similar reasoning quality.

The Anthropic cost is slightly higher per token but the runs are infrequent and short.

What the Agents Get Right

The Content Manager is reliably useful. Give it a keyword strategy and a voice document, and the draft quality is good enough that I edit rather than rewrite. This saves 2–3 hours per post compared to drafting from scratch.

The CMO catches follow-ups I would miss. When you are context-switching between four projects, a client interaction from 12 days ago is genuinely easy to forget. The CMO does not forget.

What the Agents Get Wrong

The CTO agent proposes technically correct but sometimes contextually wrong improvements. It sees the code but not the reason I wrote it that way. I reject about 40% of its proposals.

The Content Manager occasionally proposes posts that do not fit the voice — too formal, or targeting a keyword that is adjacent to but not actually in the content strategy. These get rejected quickly. The quality of the system prompt is the primary variable here.

The Ceiling

For solo developers and consultants, this agent setup covers a surprising amount of the content and CRM overhead. The ceiling is real tasks that require genuine judgment — pricing a project, handling a difficult client conversation, deciding whether to pivot a content strategy. Those stay human.

---

Resources

• —Groq API console — sign up and get an API key; the free tier is generous enough for a personal agent system
• —DeepSeek API documentation — pricing, rate limits, and the reasoning model API
• —Anthropic API — tool use guide — the reference for building reliable tool-use loops with Claude

Related Posts

• —API Key Management: Building a Central Vault for Multiple AI Providers — how all three providers' keys are managed from one admin page
• —Groq vs. DeepSeek vs. Llama-3.3-70B: Why I Mix Models Instead of Picking One — the routing logic behind which model handles which agent task
• —Case Study: santoshpaudel.me — A Portfolio That Runs Like a Business — the full platform context that the agent system runs inside

---

Want an AI content system for your own site or business? I build autonomous agent pipelines on top of Supabase and Next.js. See my services or get in touch.