A year ago, producing 300 pieces of content per week for a small company would have required a full marketing department — writers, editors, designers, a social media manager, a newsletter coordinator, and a project manager to keep them all in sync. Today, a two-person founding team at KOAT runs an AI agent workflow that does roughly the same volume before breakfast.

This is not a story about replacing humans with automation. It is a story about how small teams can now operate at a scale that was previously impossible without significant headcount — and what the real challenges look like when you build something like this in production.

The Setup: What "AI Agent Team" Actually Means

When most people hear "AI agent," they imagine a single chatbot answering questions. What we built is different: a multi-agent system where specialized AI agents handle specific roles in a content production pipeline, passing work between them the way a real editorial team would.

The team structure mirrors a human marketing organization. A research agent scans trends and competitive signals each morning. A strategy agent translates that intelligence into a content plan for the day — which topics, which formats, which platforms. A writing agent produces first drafts across multiple content formats. A review agent checks quality and brand consistency. A publishing agent formats, schedules, and posts to each channel. A reporting agent closes the loop each evening with performance data.

Each agent has a defined role, specific tools it can access, and a clear output format. The agents communicate through structured handoffs — one agent's output becomes the next agent's input, with the human team reviewing at key decision points rather than every individual piece.

The Numbers: What 300 Pieces Actually Looks Like

In a typical week, the system produces approximately: 10 long-form blog posts (1,000-1,500 words each, bilingual Korean and English); 70 Threads posts across two languages; 5 newsletter segments; 14 YouTube video scripts; 30 card news image briefs; 20 Tistory posts; and roughly 150 shorter social snippets adapted from longer pieces. The exact count varies by week based on trending topics and campaign priorities.

The quality is not uniform across all 300 pieces. The long-form blog posts receive the most attention from both the AI agents and the human review step. The shorter social content — particularly the Threads posts — runs largely automated with a lighter human-in-the-loop check. The ratio of human review time to content produced is roughly 2-3 hours of human attention per day to maintain the full pipeline.

The Technical Stack

The entire system runs on Claude (Anthropic's large language model) as the core intelligence layer, accessed through Claude Code — Anthropic's CLI interface that enables agentic, multi-step workflows. Claude Code allows agents to read files, execute scripts, write new files, and chain operations together in ways that go far beyond simple prompt-and-response interactions.

Each agent is configured through a combination of system prompts (defining role, tone, constraints) and slash commands (defining the specific workflow to execute). The content pipeline is orchestrated through a series of these commands that can be run sequentially or in parallel depending on dependencies. Research and intelligence gathering runs in parallel; writing cannot start until the strategy step is complete; publishing cannot run until quality review passes.

Supporting infrastructure includes Notion as the content database (tracking status of every piece from draft to published), Gmail for newsletter distribution, and platform APIs for automated posting to Threads and YouTube. The website's blog system is a static HTML site deployed via Netlify — simple and fast, with no CMS overhead that could introduce failure points.

What Worked Better Than Expected

The bilingual content pipeline was the most pleasant surprise. Producing both Korean and English versions of every piece doubles the content volume without doubling the time. The AI handles translation-plus-localization — not just literal translation, but adjusting examples, cultural references, and tone to fit each language's native reading patterns. A finance article written for a Korean audience sounds different from the same article written for an international English audience, even when covering the same topic.

The trend-to-content pipeline speed was also better than expected. When a significant economic event breaks — a major central bank decision, a geopolitical development, a tech product launch — the research agent can identify its content relevance, the writing agent can produce an analysis piece, and the publishing agent can have it live within 2-3 hours. A human team doing this well would take 6-8 hours minimum. The ability to move at market speed on content is a genuine competitive advantage for a small publisher.

The consistency of brand voice was also stronger than anticipated. Human writers, even good ones, have off days, varying moods, and inconsistent adherence to style guides. The AI agents apply brand voice guidelines with perfect consistency — not because the output is robotic, but because the system prompt constraints are applied uniformly to every piece. When we update the style guide, all future content reflects the change immediately.

What Was Harder Than Expected

Platform safety rules proved to be the most painful operational lesson. Early in the system's deployment, the Facebook posting agent was running at too high a frequency — posting 10 pieces in a 50-minute window. The platform's automated spam detection flagged the account, and the account was permanently deactivated. No appeal, no recovery. Months of audience-building gone in an afternoon.

The lesson: AI agents are capable of operating at speeds that human intuition would never attempt — and platforms have hard limits on posting velocity that cannot be exceeded without consequences. Every automated posting agent now runs with mandatory cool-down timers, platform-specific daily limits, and circuit-breakers that halt the pipeline if rate limit warnings appear. The system must be slower than it technically could be, as a deliberate safety constraint.

URL integrity was a second hard lesson. Early versions of the content pipeline occasionally generated plausible-looking but incorrect app store URLs — hallucinated strings that looked like valid links but pointed nowhere. An incorrect URL in a published post advertising an app is worse than no URL — it creates a broken experience for the reader and a credibility problem for the brand. All content now runs through a URL validation step before publishing, cross-referencing a curated file of verified app store links. Any URL not in that file is blocked.

The Human Role in an Agentic System

Running a high-volume AI agent content system does not eliminate the human role — it transforms it. The human team's job shifts from producing content to designing the system that produces content, monitoring the output quality, handling edge cases the agents cannot resolve, and making strategic decisions about direction and priorities.

This is a meaningful skill shift. The ability to write well still matters, but it now operates at the level of writing good system prompts and reviewing outputs rather than writing every piece from scratch. The ability to think in workflows — to decompose a content goal into a sequence of agent tasks with clear inputs and outputs — becomes a core competency. And the ability to identify when the system is producing subtly wrong outputs (factually correct but off-brand, or grammatically clean but culturally tone-deaf) requires genuine expertise that cannot be delegated back to the AI.

The best AI systems amplify human judgment rather than replace it. The human's job becomes knowing what good looks like — which is harder, not easier, than doing it yourself.

What Comes Next

The next evolution is tighter feedback loops. Currently, the system produces content and publishes it, but the performance data — which posts generated engagement, which drove app downloads, which topics resonated with which audiences — feeds back into the next week's strategy only through a manual review step. Building automated performance feedback directly into the agent pipeline would allow the system to learn and improve its content strategy with each passing week, not just maintain it.

The potential of agentic AI for content operations is still in its early stages. Teams willing to invest in building and refining these systems now will have a significant structural advantage over those who wait for the technology to become more accessible and obvious. The window where the learning curve represents a moat, rather than just a cost, is open — but it will not stay open indefinitely.