Agentic Engineering
Engineering with AI coding agentsThe essays, talks, and guides that defined agentic engineering — the discipline of building real software with AI coding agents, one step past vibe coding.
“From a viral shorthand to an engineering discipline: how 'accept the diff and hope' became patterns, harnesses, and orchestrated agent fleets in under two years.”
What is Agentic Engineering?
Agentic engineering is the discipline of building production-grade software with AI coding agents — the serious practice that forms one step past "vibe coding." Where vibe coding means accepting whatever the model produces and hoping it works, agentic engineering keeps a human firmly in charge of the parts that matter: writing the spec, running evals, reviewing the output, and owning the result.
Andrej Karpathy popularized "vibe coding" in early 2025; Simon Willison drew the line to "agentic engineering" as the serious discipline built on top, and labs like Anthropic sharpened it into patterns, harnesses, and workflows through 2026.
Autocomplete
Finishes your line
Copilot
Suggests blocks; you accept
Vibe coding
Accept the diff, hope it works
Agentic engineering
Delegate the labor — keep the spec, evals & verdict
Autonomous agents
The agent decides and ships alone
The discipline isn't "more AI" — it's high delegation with human judgment still in the loop. You can outsource the typing, not the understanding.
About this vault
A curated map of how 'vibe coding' became 'agentic engineering.' The trail runs from Karpathy's original February 2025 post and Willison's line-drawing response, through the demystification essays (Thorsten Ball's 300-line agent) and the vendor playbooks (Anthropic's workflow-vs-agent vocabulary, OpenAI's orchestration guide), to the 2026 professionalization: Willison's patterns guide, Böckeler's harness engineering, and Yegge's agent-fleet maximalism. Selection standard: primary sources and first-hand practitioner writing only — no explainers of explainers. The deepest running disagreement to watch: Yegge argues the future is orchestrating colonies of agents; Willison argues the craft is disciplined single-agent practice with tests and review. Both are in here; read them against each other.
The mainstream view
Across the primary sources, a rough consensus has formed on what agentic engineering actually asks of you:
You can outsource thinking, but not understanding.
Karpathy's line: shipping agent-written code you don't understand just defers the cost to the first production incident.
Separate "workflows" from "agents" — and prefer the simplest thing that works.
Anthropic's widely-cited framing: most production value comes from composable workflow patterns, not fully autonomous agents. Add complexity only when it earns its keep.
Not all AI-assisted coding is vibe coding.
Willison's line-drawing: reviewing, testing, and understanding the code is the professional practice; vibe coding is specifically the mode where you skip that.
By 2026 the craft is patterns and harnesses, not prompts.
The discipline now has its own literature: reusable patterns, and "harness engineering" — the test and review scaffolding that makes agent output trustworthy.
Open questions the field is still working out
Where thoughtful practitioners still disagree:
Do you actually need an agent framework?
One camp says a good framework's real payoff is cheaply swapping models and getting observability for free. Engineers who have shipped at scale push back: the bottleneck is model behavior, not the API surface, and frameworks often add more obscurity than they remove. Rough middle ground — start without one, add it only when evals or observability justify it.
Fleets or craft?
Steve Yegge argues the future is orchestrating colonies of agents in parallel; Simon Willison argues the durable skill is disciplined single-agent practice with tests and review. Both are in the reading list below — read them against each other.
What must a human never hand off?
A recurring theme in mid-2026 discussion (e.g. around Addy Osmani's essay): agents can own the inner loop, but humans must keep the outer loop — quality, evidence, and the final verdict — and guard against "cognitive surrender," rubber-stamping output they don't understand.
Frequently asked questions
What is agentic engineering?
Agentic engineering is the discipline of building production-grade software with AI coding agents — the serious practice one step past "vibe coding." The agent does the labor (reading files, writing code, running tests, iterating), while the engineer sets the spec, reviews the output, runs evals, and owns the result. The skill shifts from writing code to specifying, verifying, and understanding it.
Who coined the term "agentic engineering"?
Andrej Karpathy popularized "vibe coding" in early 2025; Simon Willison drew the line to "agentic engineering" as the professional discipline built on top of it in late 2025. Anthropic, OpenAI, and practitioners like Steve Yegge and Birgitta Böckeler sharpened it into patterns, harnesses, and workflows through 2026.
Agentic engineering vs vibe coding — what's the difference?
Vibe coding means accepting whatever the model produces and hoping it works — you don't necessarily read or understand the code. Agentic engineering keeps a human in charge of the parts that matter: the spec, the tests and evals, the review, and the final verdict. Same tools, opposite discipline: vibe coding blurs the line between intention and understanding; agentic engineering insists on it.
What should I read first?
Start with the three anchor pieces in the "Start Here" section below: Karpathy's talk that named the shift, Anthropic's "Building Effective AI Agents" (which defines workflows vs agents), and Simon Willison's essay drawing the line between AI-assisted coding and vibe coding. Then follow the timeline from the 2024 foundations to the 2026 professionalization.
Do you need an agent framework to do agentic engineering?
It's contested. One camp values frameworks for cheap model-swapping and built-in observability; engineers who have shipped at scale counter that the bottleneck is model behavior, not the API surface, and that frameworks often add obscurity. The rough consensus: start without a framework and add one only when your evals or observability needs justify it.
15 articles
Start Here
Three pieces that define the term and the debate: the talk that named it, the guide that defined it, and the essay that drew the original line.
Andrej Karpathy: From Vibe Coding to Agentic Engineering
Karpathy revisits his own 'vibe coding' coinage a year later and argues that agentic engineering is the serious discipline forming on top of it: you can outsource the typing and even much of the thinking, but never the understanding.
Agentic Engineering Patterns
An evolving, chaptered guide to the coding practices that get the best results from agents: red/green TDD, linear walkthroughs, hoarding reusable playbooks — plus the field's most-cited working definition of the term itself.
Not all AI-assisted programming is vibe coding (but vibe coding rocks)
The essay that drew the line: vibe coding — accepting output unreviewed — is fine for throwaway projects, but professional software still requires review, tests, and comprehension. The conceptual seed of agentic engineering.
The Practice: Patterns, Harnesses, Context
The working engineer's shelf: how agents actually work, what feeds them, and the harness that makes their output trustworthy.
How to Build an Agent (or: The Emperor Has No Clothes)
A working code-editing agent in about 300 lines of Go with three tools — demonstrating that the core loop behind coding agents is genuinely simple, and the sophistication lives in the harness.
Harness engineering for coding agent users
Agent = Model + Harness. Böckeler's framework separates guides (feedforward controls that prevent problems) from sensors (feedback controls that catch them) as the path to trusting agent-generated code.
Building Effective AI Agents
The essay that gave the field its working vocabulary: workflows (predefined LLM+tool paths) versus agents (LLMs that direct their own process), and the case for the simplest pattern that solves the problem.
Effective Context Engineering for AI Agents
The essay that mainstreamed 'context engineering': treating the agent's full token budget — system prompt, tools, retrieved data, history — as a finite resource with diminishing returns, managed via compaction and just-in-time retrieval.
Agentic Coding: The Future of Software Development with Agents
Flask creator Armin Ronacher's field notes on agent-driven development: which languages and tool choices reduce agent confusion, how to manage context, and real examples from CI debugging to workflow automation.
A Practical Guide to Building Agents
OpenAI's 34-page guide distilled from enterprise deployments: when an agent is actually warranted, single- versus multi-agent orchestration patterns, and guardrail design for safe deployment.
The Debate: Fleets vs Craft
Yegge's agent-colony maximalism versus Willison's disciplined craft — the live disagreement that will define the next two years.
The Future of Coding Agents
Yegge argues development is shifting from one powerful agent to orchestrated colonies of coding agents, and predicts small teams running agent fleets will outpace much larger organizations.
Vibe Coding: Building Production-Grade Software With GenAI, Chat, Agents, and Beyond
The first book-length treatment of building real software with AI agents. Kim and Yegge introduce the FAAFO framework — Fast, Ambitious, Autonomous, Fun, Optionality — for professional-grade AI-assisted development.
From IDEs to AI Agents with Steve Yegge
Yegge lays out his levels-of-adoption framework — from ignoring AI to running parallel agent fleets — and argues monolithic codebases are the biggest blocker to enterprise agent adoption.
My fireside chat about agentic engineering at the Pragmatic Summit
Willison and Statsig's Eric Lui walk through the phases developers actually go through when adopting AI coding tools, why sandboxing coding agents matters, and how test-driven development changes when an agent writes the tests.
Origins & Context
The primary sources: the post that coined vibe coding and the Software 3.0 frame everything else builds on.
There's a new kind of coding I call “vibe coding” (original post)
The origin post that coined 'vibe coding' — fully giving in to the vibes, forgetting the code exists — and the etymological root of the entire term family, agentic engineering included.
Andrej Karpathy: Software Is Changing (Again)
The Software 3.0 talk: hand-written code (1.0), learned weights (2.0), natural-language prompts (3.0) — and the prediction of a coming 'decade of agents.'
What the community is discussing right now
July 2026The conversation has moved past "what is agentic engineering" to what discipline and tooling it actually demands — and what humans must not hand off. A snapshot of the live discussion:
“Karpathy defined agentic engineering as the discipline that separates production agent work from vibe coding. The core skills he listed were spec design, eval loops, and security oversight.”
@_avichawla · X · 416 likes · Jul 2026 — the field is professionalizing
“the ability to swap out APIs just isn't the bottleneck.. like ever. It is always the behavioral issues or capability differences between models. The frameworks just usually add more complexity, obscurity, and API misalignment.”
XenophileJKO · Hacker News · on "Building Effective AI Agents" — skeptical of frameworks
“This article remains one of the better pieces on this topic, especially since it clearly defines which definition of "AI agents" they are using at the start!”
simonw (Simon Willison) · Hacker News · on "Building Effective AI Agents" — praise
Two tensions run through the mid-2026 discussion. First, tooling: the push toward unified agent stacks (Google's ADK/A2A, which drove avichawla's thread) drew immediate pushback that generic eval tooling rarely survives contact with a real product — the bottleneck is model behavior, not the API surface. Second, craft: a widely-shared thread summarizing Addy Osmani's essay argued humans must keep the "outer loop" — quality, evidence, verdict — and warned against "cognitive surrender," blindly accepting agent output. The through-line: the field agrees agents can do the labor; it is still arguing over what discipline keeps that labor trustworthy.
Start reading, not hoarding.
Import this vault to Burn 451 and actually read what matters.
New to Burn? See how the read-later app works →
Content attributed to original authors. Burn 451 curates publicly available writing as a reading index. For removal requests, contact @hawking520.