Stop Vibe Coding, Start Agentic Engineering – Micky

Key Concepts

• Agentic Engineering: A paradigm shift where developers use AI agents in parallel to execute tasks, moving beyond simple prompt-based interactions.
• Harness: The wrapper around an AI model (including API tools, system prompts, and markdown files) that guides the model to perform specialized actions.
• Context Engineering: The practice of providing the AI with precise, minimal, and relevant information (codebases, specific snippets) to prevent the model from becoming "dumb" due to context bloat.
• Service Layer: A structural pattern where code is organized into reusable functions to prevent the AI from rewriting redundant code and to reduce "code smell."
• Agentic Loop: A workflow where an agent iteratively performs a task, receives feedback (e.g., via confidence scores), and self-corrects until a desired outcome is achieved.

---

1. The Modern AI Development Stack

Mickey, a senior developer, emphasizes that 95% of his current code is AI-generated. He argues that the "vibe coding" era is over, replaced by a serious, methodical approach to engineering.

• Harnesses: He identifies Cursor as the superior harness due to its ability to switch between models and its advanced agentic view.
• Models: He currently utilizes GPT-5.5 Extra High for complex, large-scale codebases and Opus 4.7 Max for UI/frontend tasks.
• The "Harness" Philosophy: A model is merely a next-token predictor. The harness provides the "brains" by supplying tools (file reading, search capabilities) that allow the model to interact with the environment.

2. Methodologies and Frameworks

Context Engineering

• Avoid Bloat: Models perform best when the context window is kept in a "sweet spot." Overloading the context makes the agent less effective.
• Code as Truth: Instead of relying on human-written documentation, Mickey uses the "Open Source" repo (by Vercel) to fetch source code directly into his project. This provides the agent with the most accurate "source of truth."
• Minimalism: He advocates for small, modular Pull Requests (PRs) rather than massive, complex tasks.

The "Service Layer" Pattern

To prevent the AI from constantly rewriting existing functions, Mickey enforces a service layer. This ensures that when a new feature is added, the agent reuses existing logic rather than creating redundant, messy code.

The "Grep Loop" (Automated Review)

Mickey uses Grepile for code reviews.

1. Confidence Scoring: The agent reviews the PR and assigns a confidence score (e.g., 3/5).
2. Automated Correction: Using the /grep loop command, the agent reads the feedback, applies fixes, and re-submits until it achieves a 5/5 score.
3. Testing: The models are instructed to write tests for every feature; the loop continues until all tests pass.

3. Real-World Applications

• Contract Negotiation: Mickey used Claude to analyze a 27-page contract, resulting in a 3x increase in his compensation by identifying unfavorable terms.
• Accounting Automation: Instead of paying thousands of dollars for manual accounting, he used an AI agent to interface with his accounting software’s API, completing the work in two hours.
• Security: He suggests a "14-day rule" for packages: never allow an agent to install a library that is less than 14 days old to avoid malicious attack vectors.

4. Key Arguments and Perspectives

• The "Delusion" Factor: Mickey argues that successful builders in places like San Francisco succeed because of their extreme belief in their product. He encourages developers to launch early, even if the product is imperfect, to gather feedback and iterate.
• Human-in-the-Loop: Despite the high level of automation, the human must remain the "architect." The AI is a "dumb person with a photographic memory"—it needs guidance, structure, and oversight to avoid shipping bad code.
• The Future of Knowledge Work: Mickey is more bullish on AI’s impact on general knowledge work than just software engineering. He believes that as tools improve, the distinction between "technical" and "non-technical" will vanish, and those who embrace these tools will gain a massive competitive advantage.

5. Notable Quotes

• "In agentic engineering, you're doing the thinking and then you're just letting your minions do the work."
• "The model doesn't think the way humans think... you have to almost treat this like a really dumb person with photographic memory."
• "Watching without action is pointless."

---

Synthesis/Conclusion

The future of engineering is not about typing prompts, but about building sophisticated harnesses that allow AI agents to operate in autonomous, self-correcting loops. By focusing on context engineering, clean code structure, and maintaining human oversight, developers can ship products at 100x the speed of traditional methods. The primary barrier to entry is no longer technical skill, but the willingness to adopt these tools, invest in high-tier models, and maintain the "grit" to iterate through failures.