Matt Pocock’s Agentic Engineering Workflow (just copy him)

Key Concepts

• Strategic vs. Tactical Programming: The distinction between high-level architectural decision-making (strategic) and day-to-day syntax/bug fixing (tactical).
• The Harness: The environment, prompts, and tools surrounding an AI model that dictate its effectiveness.
• AX (Agent Experience): The quality of a codebase and environment that allows an AI agent to operate efficiently, analogous to DX (Developer Experience).
• AFK (Away From Keyboard) Agents: The practice of delegating tasks to autonomous agents that operate independently of the user.
• Stateful vs. Stateless Skills: Skills that maintain local context/memory (stateful) versus those that are simple, one-off procedures (stateless).
• The Bitter Lesson: An ML concept suggesting that raw compute and model scaling eventually outperform human-engineered optimizations.

---

1. Strategic Programming in the Age of AI

The speaker argues that AI has effectively "eaten" tactical programming—the manual writing of code and syntax debugging. Because AI can perform these tasks cheaper and faster, the primary value of a developer has shifted to strategic programming.

• Core Responsibility: Developers must act as "generals," focusing on codebase architecture, interface design, and velocity.
• The Multiplier Effect: AI acts as a force multiplier. A senior developer’s domain expertise serves as the "ceiling" for what the AI can achieve; if the human’s skills are low, the AI’s output is limited.
• Delegation: Effective delegation requires well-scoped tasks, clear interfaces, and robust test suites.

2. The "Harness" Over the "Model"

A central argument is that users are too obsessed with the "shiny new model" (the engine) rather than the "harness" (the chassis and aerodynamics).

• Control: Developers have significantly more control over the harness (prompts, environment, skills) than the model itself.
• Token Optimization: Instead of trying to optimize for token spend via prompt engineering, developers should focus on creating a clean, modular codebase. A well-architected codebase allows even "stupider" (cheaper) models to perform effectively because the guardrails are already in place.
• Fundamentals: The speaker emphasizes that 30–40-year-old software engineering principles remain the most effective way to ensure long-term success, regardless of which AI model is currently trending.

3. The "Teach" Skill and Personalized Education

The speaker introduced a "Teach" skill designed to fill knowledge gaps for "vibe coders" (those who can read code but lack deep engineering fundamentals).

• Methodology: It uses educational principles like the Zone of Proximal Development and Knowledge/Skills/Wisdom frameworks.
• Stateful Learning: Unlike simple prompts, this skill is stateful; it saves the user's mission, learning history, and progress locally in the workspace.
• Implementation: It generates rich HTML-based lessons and quizzes to increase "storage strength" of information, moving beyond simple terminal-based interaction.

4. Agentic Engineering: Procedures vs. Abilities

The speaker categorizes AI skills into two types:

• Procedures: Skills invoked by the human to control the AI (e.g., the "Grill Me" skill, which turns an agent into an adversarial interviewer to stress-test ideas). The speaker prefers this approach to maintain human control.
• Abilities: Skills the model invokes itself. The speaker warns that leaking too many ability descriptions into the context window can degrade performance.

5. AFK Agents and Queue-Based Workflows

The speaker advocates for moving away from "human-in-the-loop" (HITL) for every task and toward AFK (Away From Keyboard) workflows.

• The Queue Model: Development should be viewed as a queue of tasks. Instead of an infinite "loop" (which the speaker views as a marketing tactic to sell tokens), developers should triage tasks, label them, and let agents resolve them independently.
• Sand Castle: A tool developed by the speaker to run agents inside secure, isolated sandboxes (Docker/Podman), allowing for parallelized agent execution without risking the host system.

6. Synthesis and Actionable Advice

The speaker concludes that the most successful AI builders will be those who:

1. Maintain Human Oversight: Use AI to handle tactical implementation, but retain control over product vision and architectural decisions.
2. Build Self-Improving Systems: Instead of relying on a model to find bugs, build automated test suites and cron jobs that proactively scan for issues.
3. Reset and Simplify: For those struggling with AI, the speaker recommends deleting all custom plugins and "bloated" instructions. Start with a blank slate, observe the agent's baseline behavior, and layer on only the necessary procedures.

Notable Quote: "Your skills are the ceiling on what AI can do. If your skills are low, then AI is not going to be able to go past that."