Agentic Engineering
Building software with coding agents that both write and execute code.
Term popularized by Simon Willison (@simonw), Late 2025.
What it is, why now
Agentic engineering is the practice of building production software using AI coding agents—tools like Claude Code, Codex CLI, Aider, and Cursor's agent mode—where the agent both generates and runs code, not just auto-completes it. The term was popularized by Simon Willison in late 2025 to distinguish this workflow from earlier "AI-assisted coding" (copilot-style autocomplete) and from "vibe coding" (natural-language one-shot apps, coined by Karpathy in Feb 2025).
The defining property: the agent completes full engineering loops autonomously — reading files, running tests, interpreting errors, iterating. This changes the economic unit of software work. One senior engineer running 3-8 agents in parallel produces output that previously required a team. The skill shifts from writing code to writing clear specifications, reviewing agent output, and building sufficient verification (tests, types, logs) for agents to self-correct.
What matters for builders now: (1) context management — agents forget, so state must be explicit and short-lived; (2) verification loops — your tests and types are now an agent-facing API, not just human documentation; (3) failure modes — agents produce bloated abstractions, aesthetic inconsistency, and subtle bugs that pass tests. The craft is learning which 80% to delegate and which 20% to intervene on.
How we got here
- Feb 2025
Karpathy coins "vibe coding"
Andrej Karpathy tweets about building apps by describing them in natural language, no code written by hand. This is the precursor aesthetic to agentic engineering — but still mostly one-shot, not iterative.
MenuGen case study - Apr 2025
Karpathy ships MenuGen
First widely-cited end-to-end vibe-coded app. Karpathy documents the "80% appears in 30 minutes, 20% takes 40 hours" pattern — the first honest accounting of what agent-built code actually costs.
- Jun 2025
Claude Code moves agent coding mainstream
Anthropic ships Claude Code as the first widely-distributed agent coding tool with good defaults. Becomes the reference implementation — Karpathy later writes "within weeks, from 80% hand-coded / 20% agent to 80% agent / 20% human tweaks."
- Oct 2025
Simon Willison coins "agentic engineering"
Simon starts his "Agentic Engineering Patterns" newsletter to distinguish iterative agent-driven workflows from earlier autocomplete/chat models. Term catches on in the AI engineering community.
Agentic Engineering Patterns newsletter - Nov 2025
The inflection point
Per Simon: "November 2025 was when AI coding agents crossed from mostly works to actually works." Codex, Claude Code, and Cursor agent mode all reach production viability in the same quarter.
- Mar 2026
Karpathy on code agents + autoresearch
No Priors podcast episode where Karpathy lays out the full picture: agents for coding, agents for research, and what breaks at scale. Introduces "agent code quality is objectively worse but productivity is objectively higher" as the honest trade-off.
- Apr 2026
Google insider claim vs rebuttal
Steve Yegge's Google insider claim: 60% of engineers still on basic chat tools, 20% are agentic power users. Demis Hassabis publicly rebuts. Illustrates the adoption is still very uneven — even inside frontier labs.
The 10 pieces that matter most
Curated from across Burn 451's vaults. Each piece has an AI summary — click to read it on its home vault page.
Vibe coding MenuGen
From the Karpathy vault
The first full public case study of building with agents end-to-end. Required reading for the 80/20 split.
Claude coding notes — 80% agent, 20% edits
From the Karpathy vault
Karpathy's timestamp on when the ratio flipped. If you started 2025 assuming agent coding was a toy, this is the data point that changed most practitioners' minds.
Agent code quality — bloated abstractions, poor aesthetics
From the Karpathy vault
The honest counter-essay. Agents work but produce ugly code. This is the 20% problem, described.
Optimal LLM-assisted coding — diversifying workflows
From the Karpathy vault
Karpathy's taxonomy: which tool for which kind of task. Autocomplete vs chat vs full agent — not always the same right answer.
8 agents in parallel — 4 Claude, 4 Codex
From the Karpathy vault
Early evidence that parallelism works. The coordination overhead is real but outweighed by raw throughput.
Agentic engineering on Lenny's Podcast
From the Simon Willison vault
Simon's clearest long-form articulation of the patterns: context management, verification loops, the "lethal trifecta".
Eight years of wanting, three months of building with AI
From the Simon Willison vault
Personal case study: software projects Simon had shelved for years, shipped in a weekend with agents. The psychological shift matters as much as the technical one.
Fireside chat on agentic engineering at Pragmatic Summit
From the Simon Willison vault
For a more structured dev audience. Covers the tradeoffs from the senior engineer POV — what to delegate, what to still own.
Vibe coding SwiftUI apps is a lot of fun
From the Simon Willison vault
Simon applies Karpathy's pattern to Swift. Less a tutorial than evidence that the agentic approach generalizes across languages.
Steve Yegge on AI adoption at Google
From the Simon Willison vault
The hard data on adoption unevenness. Even at Google, most engineers are not yet running agents daily.
Related concepts
Vibe Coding
Adjacent concept. Vibe coding = natural-language one-shot apps; agentic engineering = full iterative software development with agents.
AI Bookmark Management
MCP makes reading lists AI-accessible — agents can search, analyze, and build on a user's curated knowledge through AI bookmark management tools.
Want to read more like this?
Burn 451 is a reading tool that helps you actually finish articles instead of hoarding them. Import a Vault, set a timer, read what matters.
Concept page curated by @hawking520 · Burn 451 · Last updated 2026-04-14