Andrej Karpathy's AI Engineering Playbook (2026)

Andrej Karpathy's AI Engineering Playbook: Every Framework in One Place

No one has shaped how builders think about working with AI more than Andrej Karpathy. He named Software 3.0, drew the line between vibe coding and agentic engineering, popularized the CLAUDE.md file, and published the knowledge-base pattern half the serious builders you follow now use. In May 2026 he joined Anthropic's pre-training team to get back to frontier LLM research, which only sharpens why his frameworks are worth studying as a set.

The problem is they are scattered across talks, gists, and X threads. This page connects them into one playbook - the worldview, the practice, the config files, and the knowledge layer - and links the full guide for each.

                 KARPATHY'S AI ENGINEERING PLAYBOOK
                                │
   ┌──────────────┬─────────────┼─────────────┬──────────────┐
   v              v             v             v              v
 WORLDVIEW     PRACTICE      PARALLELISM    CONFIG         KNOWLEDGE
 Software 3.0  Agentic Eng   Parallel       CLAUDE.md +    LLM Wiki
               (spec/diff/   agents like a  agents.md      (compounding
 "context      eval loops)   tech lead      (the rules)    knowledge)
  window
  is code"

1. The worldview: Software 3.0

Start here, because it reframes everything else. Karpathy's argument is that the context window is now your program and the LLM is the interpreter - you program in English, not Python. That single shift is why prompt files, specs, and knowledge bases suddenly matter more than syntax.

Read the full breakdown: Karpathy's Software 3.0: The Context Window Is Code - what it means, and what kinds of software should stop existing.

2. The practice: agentic engineering

At Sequoia Ascent 2026, Karpathy drew a hard line between vibe coding (prompt and pray) and agentic engineering (spec design, diff review, eval loops). The difference is discipline: you direct the agent like an engineer, not a slot machine. This is the skill that actually separates a 10x builder from someone fighting their tools.

Read the full framework: Agentic Engineering: Karpathy's New Framework - spec design, diff review, and eval loops, explained.

3. Parallelism: run agents like a tech lead

Once you direct agents instead of typing for them, you can run several at once. Karpathy manages multiple Claude Code sessions like a small engineering team - one window per feature, reviewing diffs and merging outputs. The mental-model shift is from coder to tech lead.

Read the exact setup: How to Run Parallel Claude Code Sessions Like Karpathy - the four-step workflow and the failure modes to watch.

(This is also where the playbook meets production reality: running many agents creates a merge-and-verify bottleneck, which is what a codebase harness and Crabbox exist to solve.)

4. The config files: CLAUDE.md and agents.md

Directing agents well means writing down the rules once. Karpathy's CLAUDE.md rules are the canonical example - a short file that, in community testing, cut Claude's coding mistakes dramatically. The same instinct generalizes to an agents.md for any agent runtime.

Read both:

• Karpathy CLAUDE.md Rules: Cut Claude Errors 41% to 11% - the exact file and a two-minute install.
• Karpathy's agents.md: What It Is and Why It Matters - the cross-runtime version of the same idea.

5. The knowledge layer: the LLM Wiki

His most recent shipped concept (April 2026). Instead of RAG retrieving raw fragments every time, an LLM incrementally builds and maintains a persistent, interlinked wiki of markdown - knowledge that compounds with every source you add. It is Software 3.0 applied to your own knowledge.

Read the full architecture: Karpathy's LLM Wiki: A Knowledge Base That Compounds - the three layers, three operations, and how to start.

How the playbook fits together

The pieces are not separate tips - they are one stack. Software 3.0 is the why. Agentic engineering is the how you work. Parallel agents is the how you scale. CLAUDE.md and agents.md are how you encode the rules. The LLM Wiki is how your knowledge compounds. Learn them as a set and you stop fighting AI tools and start directing them.

Go deeper with our courses:

• AI Agent 101 Course - turn these frameworks into agents you actually ship, with eval loops built in
• Claude Code 101 Course - CLAUDE.md setup, subagents, and the diff-review workflow Karpathy runs

If you want to put this whole playbook into practice alongside builders doing the same, join AI Builder Club.

Frequently Asked Questions

What is Andrej Karpathy known for?

Karpathy is an OpenAI co-founder and former Tesla AI director who has become the most-cited voice on how to build with AI. He named Software 3.0, defined agentic engineering, popularized the CLAUDE.md config file, and published the LLM Wiki knowledge-base pattern. In May 2026 he joined Anthropic's pre-training team.

What is the difference between vibe coding and agentic engineering?

Vibe coding is prompting an agent loosely and hoping for the best. Agentic engineering, Karpathy's term, is directing the agent with discipline - writing specs, reviewing diffs, and running eval loops. The first is a slot machine; the second is how you get reliable, reviewable output at scale.

What is Karpathy's Software 3.0?

Software 3.0 is Karpathy's framing that the context window is now the program and the LLM is the interpreter - you program in natural language instead of code. It reframes prompts, specs, and knowledge bases as the new source code, and is the worldview underneath his other frameworks.

What is the LLM Wiki?

The LLM Wiki is a pattern Karpathy published in April 2026 where, instead of RAG, an LLM incrementally builds and maintains a persistent, interlinked markdown wiki from your sources. Knowledge compounds over time rather than being re-retrieved per query. It is his most recent shipped concept.

Did Karpathy join Anthropic?

Yes. On May 19, 2026, Karpathy announced he joined Anthropic's pre-training team under Nick Joseph to get back into frontier LLM research, saying the next few years at the frontier of LLMs will be especially formative.

Related Content

• Karpathy's Software 3.0: The Context Window Is Code - The worldview.
• Agentic Engineering: Karpathy's New Framework - The practice.
• How to Run Parallel Claude Code Sessions Like Karpathy - The scaling move.
• Karpathy CLAUDE.md Rules: Cut Claude Errors 41% to 11% - The config file.
• Karpathy's agents.md: What It Is and Why It Matters - The cross-runtime config.
• Karpathy's LLM Wiki: A Knowledge Base That Compounds - The knowledge layer.

Start Here

Read them in order: Software 3.0 for the worldview, agentic engineering for the practice, then install Karpathy's CLAUDE.md rules in a real repo this week. That single file is the fastest way to feel the difference between fighting an agent and directing one.

To put the full playbook into practice with builders doing the same, join the AI Builder Club.

Join AI Builder Club