Workshop Index
Workshop Chapters
Arbeitsübersicht der Kapitel und Unterthemen aus D:\course\chapters.
Chapter 1: Foundations — What's Actually Happening
- The Agentic Loop
- What is an LLM
- Tokens — input vs output
- Pricing — API vs subscription
- Model vs Harness
- Agents are just 4-5 tools in a loop
- 12-Factor Agents — key principles
- The two levers that actually matter — fewer steps OR more accurate step selection
Chapter 2: The Landscape — Where Are We and Where Is This Going
- Stages of Agentic Engineering
- Vibe Coding vs Agentic Engineering
- Danger of CEOs — AI is imperfect but very capable
- Models will change — your moat is evals, harness engineering, and mental models
Chapter 3: Getting Your Hands Dirty — Setup and First Loop
- Prerequisites — Node.js, GitHub account, API key
- /setup page on Sievering
- Give them a CLI to test
- Git repos — clone vs own
- Why CLI over GUI
- Permissions in Claude
- The first Ralph Loop
Chapter 4: Context Is Everything — The Core Skill
- Context Window
- Context Rot
- The Dumb Zone
- Plan Mode
- Claude.md / AGENTS.md
- MD files
- Skills
- .env
- Artifact Files
- Frequent Intentional Compaction
- RPI framework and QRSPI evolution
- Human Leverage Pyramid
- "Don't read the plans, read the code"
- Instruction budget
Chapter 5: Quality and Control — Making It Reliable
- Backpressure
- Tests
- TDD for Agents
- Subagents
- Browser Use
- Prompt Injection
- "Control flow via prompt" anti-pattern
- Consistency vs Variance tradeoff
Chapter 6: Scaling Up — Team and Architecture
- Orchestrator Pattern
- Remote Control
- Reusable Skills
- Team Sharing
- CLI Loop
- Library Meta-Skill
- Background Agent Patterns
- Coaching Loop
Chapter 7: Harnesses, Models, and Providers — Choosing the Right Stack
- The stack map: application → harness → model → provider → hosting
- Harness landscape
- Why the harness matters
- Model landscape
- Provider landscape
- One model, many providers
- Data protection and procurement basics
- EU-sensitive setups
- Enterprise routes and self-hosted options
- Trade-offs: compliance, latency, quality, cost, lock-in, observability, operational burden
- Future-proofing with provider/model abstraction
Chapter 8: Codebase Poisoning — Why Messy Systems Teach Agents Bad Habits
- What codebase poisoning means
- Similarity to context poisoning
- Brownfield reality
- The danger of a dirty core
- Good code is contagious, bad code is contagious too
- Clean starts, protected cores, and golden paths
- Refactor-before-scale
- Scaffold the patterns you want copied
- Constrain generation and use the harness to fight poisoning
- Detect poisoning early
- Brownfield strategy: isolate a clean seam
Chapter 9: Autoresearch — Agents Running ML Experiments, Not Just Writing Code
- What Karpathy means by autoresearch
- The core loop: modify code → run training → measure → compare → iterate
- Minimal surface area
program.mdas human-programmed research organization- Fixed wall-clock budgets and objective metrics
- The research harness pattern
- Human role shift
- Failure modes and operational risks
- Generalization beyond ML
Chapter 10: Vocabulary — Why Words Matter When Working With Agents
- Why vocabulary is not pedantry
- The communication problem in AI
- LLM vs agent vs tool vs harness
- Vocabulary as debugging, design, and collaboration tool
- Why vague language leads to vague requests
- Domain vocabulary matters too
- Building a personal glossary
- Teaching agents your vocabulary
- Key distinctions: AI, LLM, Model, Agent, Tool, Harness, Context, Prompt, Eval
Agentic OS — Folder Structure as Agent Architecture
- The pattern: a shared workspace folder is the agent's brain
- Why it beats hosted / framework stacks
- The four layers: Rules, Context, Skills, Memory / Learnings
- Workspace structure for business use
- Agent Runner: Pi
- The core insight: folder structure is orchestration