My Job Is Literally Talking to AI Agents All Day Now

Key Concepts

• AI Agents: Autonomous systems capable of executing tasks, making decisions, and using tools to achieve a goal with minimal human intervention.
• Agentic Work: A workflow where the human acts as a manager, providing instructions and context, while the AI handles the execution, problem-solving, and tool integration.
• Claude Code: An agentic framework/tool that allows users to interact with AI models (like Claude Opus 3.5) to perform complex tasks, write code, and interact with external APIs.
• MCP (Model Context Protocol): A standardized protocol that acts as a "bridge" or "documentation" for agents, allowing them to easily connect to and utilize external tools, APIs, and data sources.
• Context Window: The limit of information an AI can "remember" during a conversation. Exceeding this limit causes the agent to lose track of earlier instructions or data.
• Vibe Coding: A term often used to describe non-technical users building software through natural language; the author argues this is misleading because it undersells the strategic and analytical power of agents.

---

1. The Evolution of AI Workflows

The author categorizes the progression of AI usage into three distinct stages:

• Old School AI: A back-and-forth interaction with a chatbot (e.g., ChatGPT/Claude web interface) to gather information, followed by the human manually using external SaaS tools to perform the work.
• AI Automations: Using platforms like Make.com or n8n to create rigid, step-by-step workflows. These are effective but lack "agentic" reasoning—they follow a pre-defined path rather than adapting to problems.
• Agentic Work: The current frontier. Using models like Claude Opus 3.5 or GPT-4.5, agents can now "think," plan, and execute tasks. They can create their own tools, debug their own code, and iterate until a mission is complete.

2. The "Manager" Workflow

The author emphasizes that he is not a coder. His role has shifted to that of a manager:

• Parallel Processing: He manages multiple agent instances simultaneously. Each instance has its own mission (e.g., SEO optimization, landing page design, market research) and operates independently.
• The Loop: The workflow follows a specific mental model: Brainstorm → Plan → Execute → Test.
• Delegation: The human provides the goal and context; the agent proposes the path, executes the code, and reports back. If an error occurs, the human asks the agent to "find a solution," and the agent iterates until successful.

3. Technical Setup and Implementation

To move from a terminal-based interface to a visual one, the author recommends:

• VS Code: Used as the primary interface to chat with Claude Code.
• Claude Code Extension: Installed via the VS Code marketplace to visualize the agent's actions.
• Initialization: Using the /init command within a project folder to define the mission.
• MCP Servers: These are critical for expanding agent capabilities. By installing specific MCP servers (e.g., for Apify or Railway), the agent gains the ability to scrape data, host instances, or interact with complex APIs without the user needing to write the underlying code.

4. Real-World Applications

The author highlights several practical use cases for agentic workflows:

• Market Research: Using Apify MCP to scrape YouTube comments or find influencers based on specific brand criteria.
• Web Development: Building pixel-perfect landing pages and CMS-like admin panels for personal websites.
• SEO Strategy: Connecting Claude Code to Google Search Console and SEO-specific MCPs to analyze data and automatically generate content.
• Infrastructure: Using Railway MCP to self-host n8n instances, significantly reducing costs compared to traditional SaaS subscriptions.

5. Limitations and Challenges

Despite the productivity gains (which the author claims are 10x higher than six months ago), there are notable hurdles:

• Context Window Constraints: With a 200,000-token limit, long-running projects with massive amounts of data can cause the agent to "forget" earlier instructions.
• "Thinking" vs. "Doing": While agents are excellent at execution, they are not always the best for deep, philosophical, or life-strategy brainstorming.
• Complexity: While the author claims it is "not that hard," the setup requires a basic understanding of project folders, extensions, and the willingness to troubleshoot when things don't work perfectly.

6. Notable Quotes

• "I’m just a manager. I give them the right task, the right instruction, the right information, and I answer the question and they walk me through all the steps until every mission is done."
• "The difference between using models and using agents is syncing versus doing."
• "It’s not all magic... there are some limitations and it’s important to talk about it even though I’m pretty sure they are going to drop significantly in the next months."

Synthesis

The shift toward agentic work represents a fundamental change in how non-technical individuals interact with technology. By leveraging frameworks like Claude Code and protocols like MCP, users can bypass traditional coding barriers to build, scrape, and analyze data at scale. The key takeaway is that the future of work is not about learning to code, but about learning to manage AI agents through clear, iterative, and structured communication.