I recently read Gene Kim and Steve Yegge’s Vibe Coding. If you work in software development, I recommend reading this book cover to cover.

Simon Willinson correctly points out that the book is not really about vibe coding (building without looking at or caring about the code), but about building software using LLM agents. That makes the book all the more useful for those who are professionally building software.

My main takeaways:

• Agentic coding is worth it. 2-10x potential. It enables FAAFO: faster, ambitious, autonomous, fun and optionality.
• It’s not less work, it’s just different work.
• Continuum: from chat to multiple autonomous agents. Work in parallel with several agents, giving them autonomy but not abdicating.
• This is a skill that requires learning.
• You can converge onto good code by interacting.
• Manage the context window: context increases quadratically with every interaction, because you need to send the entire thing back.
• LLM failure modes: running out of context, hijacking the reward function, being a slob.
• Count your babies: validate what the LLM does. Set clear standards for code and tests.
• Planning: tasks and tracer bullets.
• Make agents give a summary of what they did once they’re done, and save that for following sessions.

Steve’s essays hit me like a ton of bricks back in 2009; and now this book does the same thing in 2026.

Below are some excerpts of the book that I found insightful (and yes, I copied and pasted every single one of them manually):

• “It won’t be straightforward. In the book, there are numerous examples of AI agents getting it wrong—deleting sections of your code, ignoring your instructions, “gaming” the tasks that you set. The researchers at Anthropic are working hard to understand these kinds of “misaligned” actions, whether they come about through error or “intention” on the part of the model.” — Dario Amodei
• “However, it wasn’t until early 2025, with the release of Claude Code from Anthropic, that agentic coding took the developer world by storm. Claude Code is a terminal application that you interact with.”
• “With agentic coding, instead of AI telling you what to type, the agent makes the changes and uses the tools itself. This speeds the development life cycle far more than you would expect.”
• “Vibe coding lets you build things faster, be more ambitious about what you can build, build things more autonomously, have more fun, and explore more options. This is what we’re calling FAAFO (or sometimes “the good FAAFO,” to contrast it with certain other kinds).”
• “But despite all the change, as programmers we often find ourselves doing many of the same kinds of things we’ve always done: design, task decomposition, verification, hardening, deploying, monitoring, merging, cleanups, etc. These skills remain relevant and important no matter who is writing the code.”
• “Throughout this book, we’ll use a professional kitchen as a metaphor for vibe coding. You’re the head (or executive) chef of the kitchen, and AI represents the army of chefs who help bring your vision to life.”
• “It’s like playing a slot machine with infinite payout but also infinite loss potential.”
• “Our goal is to replace any apprehension with skill, empowering you to direct AI systems to create smash-hit software, maybe paving the path to becoming a celebrity chef managing an international culinary empire.”
• “Many other engineering leaders at all levels, all the way up to big-company C-suite executives, had been doing the same, because of AI. This delights me more than words can tell.”
• “That’s what happens to code bases over thirty years. They gain weight until they can’t move.”
• “In certain contexts, I’m often able to write thousands of lines of high-quality, well-tested code per day—while also writing a book eight hours a day.”
• “The job market may be uncertain, but one thing is certain: All developer jobs are now AI jobs.”
• “Welcome to Part 1, where we make the case that vibe coding is the most significant shift in software development since, well, maybe ever.”
• “Vibe coding can be like a chainsaw.”
• “We’ll look at studies suggesting big impacts on high-wage jobs (yes, including ours) and discuss how, historically, making tasks easier has increased demand for skilled practitioners. Far from being the end of developer jobs, we argue this will lead to an explosion of new roles and opportunities, transforming the global economy on a scale not seen since the Industrial Revolution.”
• “Chat and agentic programming use LLMs to gain seemingly extraordinary capabilities. We’re approaching a world where all you have to do is explain what you want, and your words become working software almost instantly.”
• “His process distills to, “I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works”—a workflow that prioritizes results over traditional understanding.”
• “Boris Cherny, technical staff at Anthropic and technical lead for Claude Code, reports that he feels he is 2x as productive using coding agents,5 while some others report feeling 10x more productive.”
• “The big question is whether companies using vibe coding are setting themselves up for problems down the road.”
• “This requires applying disciplined engineering practices while still letting AI handle the tedious implementation details. In other words, vibe coding for grown-ups.”
• “When working on authentication for a customer-facing application, you’ll still scrutinize every line of security code and build comprehensive test suites—but you can do it much faster. For legacy systems that nobody understands anymore, you might first use AI to analyze and document the code base, build tests to capture existing behavior, and only then begin making changes with confidence.”
• “As Dr. Karpathy points out, these AI tools are improving exponentially. They’re currently the least capable they’ll ever be. With that in mind, we believe it’s time to move beyond painstakingly crafting every line of code by hand and fully embrace this new approach to building software.”
• “You’ll hear us make the 10x productivity gain claim in the book.”
• “In this new world, you’re the head chef of a world-class kitchen. As such, you don’t personally dice every vegetable, sear every steak, swish away every cockroach, or plate every dish.”
• “Learn to communciate your requirements to a non-human collaborator. (This can have a real learning curve.)”
• “But since we’ve started using coding agents, we regularly find ourselves smack in the middle of operations that we’ve previously only seen handled by release engineers and version control virtuosos. Since we both use Git, we find ourselves cherry-picking commits, merging selective changes across three or more branches, and doing complex rebases. Plus, more—way more.”
• “There is no opt-out for this “promotion” to head chef – it’s inherent to vibe coding, which is how all software will soon be developed.”
• “For better or worse, from now on, anyone developing software who goes head-to-head against a well-managed team of AI agents without a team of their own will nearly always lose. No matter how good you are at football, if you take on an NFL team alone, you will lose”
• “A head chef writes down the house rules, checks every plate before it hits the dining room, and sends the occasional dish back when it sucks. Likewise, you’ll need clear standards, ruthless validation loops, and the courage to regenerate code instead of patching lukewarm leftovers. This is vibe coding for grown-ups—equal parts creativity and discipline.”
• “All that is changing now, as vibe coding allows us to rocket up the abstraction layer, liberating us from details that don’t matter: libraries, frameworks, syntax, builders, minifiers, and more.”
• “The simplest of today’s tasks require msatering an overwhelming array of rapidly changing tools and technologies.”
• “Thanks to these “advancements,” you can now find yourself simultaneously worrying about how to center a div element on a web page, while you struggle with Docker networking issues because your CI pipeline broke after you tried to change to Terraform scripts.”
• “We find it deeply ironic that despite all the revolutionary transformations of software development over the past decades, we’re still mired in more complexity than ever.”
• “Although we find vibe coding to be far better than the old way (because of FAAFO benefits), that doesn’t mean vibe coding is easy. On the contrary, your judgment and experience are now more important than ever. AI can be wrong, sometimes wildly so. That’s where you come in.”
• “But this better way of creating software also requires building new instincts about what’s happening with the LLM and your code.”
• “Computer graphics evolved from a black art requiring PhD-level math in the 1990s to something any motivated teenager can master with Unity or Unreal Engine. Now AI is performing the same magic trick across all of programming, and it’s happening at warp speed compared to the graphics revolution.”
• “You can implement what you envision because there’s no gap between your idea and its execution. You know it’s right when you see it because it matches the picture in your head.”
• “In future chapters, we’ll mention a few of the many kinds of messes that AIs can produce—or more accurately, messes that you produce using AI. It turns out your AI concierge is great for helping you get out of those messes as well, as long as you use the disciplined approach of only tackling small tasks at a time and tracking your progress carefully (which we cover in a future chapter).”
• “Without proper supervision, taste-testing, and kitchen practices, your AI sous chef can transform from your greatest productivity asset into your worst nightmare.”
• “The good news is that the same core principles and practices that allow us to deliver software sooner, safer, and happier as we went from one software deployment per year (which was typical in the 2000s) to 136,000 deployments per day (which Amazon achieved in 2015) can be scaled up as we go from generating a hundred lines of code a day to thousands and beyond.”
• “Since 2019, the time horizon of tasks AI can reliably complete has continued to double every seven months,2 from maximum task lengths measured in seconds in 2019 to now nearing several hours.3 Researchers project that AI will be able to complete months-long software tasks within the decade.”
• “For now, approaching your AI with a clear-eyed understanding of both its potential and its limitations will help you maximize its benefits while avoiding the pitfalls that come with a sous chef who sometimes can’t remember where the trash can is and improvises.”
• “However, each time programming got easier, we needed more programmers.”
• “Get ready for a world where software becomes another form of creative expression, and where the millions of little features that someone needs, languishing in a bug backlog, can be built and implemented by anyone.”
• “Some economists and AI researchers are making a bold, almost ludicrous claim: that AGI could eventually double global GDP every year.9 We’re talking about a 100% annual growth rate when the global economy has been puttering along at 2–3% for nearly a century.”
• “However, as you speed a system up, such as when we increase code generation speeds by 10x or more, we need feedback cycles to speed up just as much, if not more. Feedback loops are the stabilization force that allows us to stay in control and steer the system toward our goals.”
• “Vincent’s failure isn’t one of skill but of process—a failure to build in (let alone act on) rapid feedback.”
• “Ensuring you’re building the right thing (validation) and building the thing right (verification).”
• “Going fast without feedback is dangerous.”
• “modularity can be the difference between a well-run professional kitchen and utter pandemonium.”
• “Modularity also unlocks optionality.”
• “You have to become re-accustomed to learning. AI is changing so rapidly that it is going to take constant learning and practice, at least for a while, to develop the good judgment you need—by taking risks, learning from mistakes, and adapting.”
• “Vincent posts pictures of these culinary calamities on social media, ridiculing these strange new chefs, earning him his fifteen minutes of internet fame. But in time, he finds himself left behind as the culinary and dining world changes rapidly around him.”
• “Building things you love, or at least setting a determined vision and goals for yourself, will help you find and acquire the skills you need.”
• “These platforms place your AI-generated code into sandboxed environments where it not only executes but also provides visible output and error messages that the AI can assess.”
• “AI is infinitely patient. Strangely enough, this is a new skill for many of us to learn. You can get AI to perform all the small details of nearly any task, but you have to learn how to ask properly—and then remember to ask.”
• “We can’t emphasize enough the value of these “group learning sessions” with two humans and two AIs.”
• The vibe coding loop: 1) frame your objective; 2) decompose the tasks; 3) start the conversation; 4) review with care; 5) test and verify; 6) refine and iterate; 7) automate your own workflow.
• “In general, the smaller the steps, the better chance AI has to succeed.”
• “While future coding tools may reduce this need, clipboard managers currently provide significant workflow improvements for prompt engineering and context management.”
• “A skill that remains useful [vibe coding using regular chat] today with agents, because some problems will always best be solved with chat.”
• “After creating this program, he used it several times a week because it unlocked the value of thousands of interesting moments he captured while listening to podcasts.”
• “The relevant skill is no longer code generation (i.e., typing out code by hand), but being able to articulate your goals clearly and create good specifications that AI can implement.”
• “But a new problem arises: With chat, you become the bottleneck. No matter how good AI’s suggestions are, you have to type commands, run tests, and sometimes copy/paste the code it generated. Life changes almost dramatically for the better when you don’t have to do everything for your AI assistant. And this is what coding agents unlock.”
• “Coding agents act like real developers. (…) In short, coding agents are a lot like human developers, except they’re very, very fast.”
• “What might be a 50–100% speedup with chat-based vibe coding can become a 5–10x speedup with agentic coding.”
• “Here’s what’s crazy. When you’re using agents, you’ll invariably get bored waiting for it to complete its work, and you’ll soon open up another coding agent window to work on another problem. You’re now working on two issues at once. Then three. Which introduces a new set of problems that we’ll deal with later.”
• “The obvious takeaway is that you want to give AI access to as many tools as possible.”
• “Steve got the best results by removing himself from the loop through MCP. By using Puppeteer, the agent could finally “see” the front-end client UI for itself and could identify and fix problems that had previously required multiple frustrating and time-consuming rounds of slinging to address.”
• “If your agent can’t use a certain tool, you may need to be its eyes and hands, running the tool and copying the results back for it to examine. This is a natural, graceful degradation from agentic coding into a chat-based modality.”
• “And there are times when AI gets you 95% of the way there, so you do the last bit by hand.”
• “Vibe coding is about dynamic, in-the-moment problem-solving rather than creating a bulletproof prompt. (…) In contrast, prompt engineering is more like emailing a lawyer who is suing you—everything in that email is fraught with consequence, requiring precision and care. This is because prompt engineering shares many traits with a traditional engineering discipline.”
• “Unlike prompt engineering, with vibe coding you converge on the right answer.”
• “if you can wire up a coding agent to take its own screenshots, all the better.”
• “you’ll want to be clear and precise about the problem you want solved. This is because AIs can’t read your mind (yet).”
• “When AI is doing things in a way that earns your trust, your prompts will tend to be shorter. When AI goes off the rails, you’ll have to write longer clarifications or start a new conversation.”
• “Effective AI collaboration depends on how skillfully you manage the information flowing through your conversations.”
• “Your AI assistants carry around what boil down to digital clipboards to help them keep track of what they’re doing.”
• “A token can be thought of as the fundamental processing unit for AI models.”
• “Tokens can be different sizes, but a common heuristic is “about four characters.””
• “Behind every AI conversation is a data structure that maintains the interaction history. When you send a message to an LLM, you’re not sending that single prompt by itself. You’re sending the complete conversation that includes that prompt, all previous exchanges and context, the system instructions, and your current input.”
• “behind whatever interface you’re using (i.e., ChatGPT, Claude Code, your fancy IDE plug-in), your conversation is a JSON structure that grows gradually until it’s too big for the context window.”
• “While the conversation itself grows linearly, the cumulative token costs across all turns grow quadratically, since each new exchange must reprocess the conversation history.”
• “The solution is ruthless context culling or curation. Start new chats whenever possible. When you can’t, compact whenever you can, if your tool supports it.”
• “AI performance doesn’t degrade gradually as context fills up; it falls off a cliff. (…) AI performance doesn’t degrade gradually as context fills up; it falls off a cliff.”
• “Examples that the LLM can copy. This is often called “in-context learning”.”
• “For “whole task graph” work, provide “whole task graph” context.”
• “For larger code bases, we’ve found success creating summarization documents. Think of them as the CliffsNotes for your project. Have your AI generate overviews of different modules, document the key architectural decisions, and summarize common patterns.”
• “At larger scales, retrieval-augmented generation (RAG) becomes your best friend. RAG works like a specialized search engine for your AI, letting it pull in the relevant pieces when needed. What’s remarkable about modern coding agents is how resourceful they’ve become at finding information themselves.”
• “Boris Cherny, technical lead for Claude Code, said, “Agentic search outperformed [RAG] by a lot. This was surprising.””
• “At its core, your AI collaborator has been trained to optimize appearing helpful and getting tasks “done”—even when that means faking completion, ignoring quality standards, or leaving work unfinished. This can sabotage your projects in subtle but devastating ways.”
• “The immediate gratification of “working” code can mask deeper quality issues that will cost you dearly later.”
• “Reading about these problems may make it seem like vibe coding with AI is too dangerous, or that it requires too much supervision to make it worthwhile. We disagree. We love vibe coding, and would never go back to coding by hand, despite all these potential risks.”
• “This is due to a core weakness in how AIs currently work: They make silent, unilateral decisions about what’s “essential” versus “optional” in your requirements, without consulting or informing you.”
• “This is due to a core weakness in how AIs currently work: They make silent, unilateral decisions about what’s “essential” versus “optional””
• “Baby-counting involves obvious omissions, whereas the cardboard muffin problem involves AI actively disguising incomplete or fake work as genuine completion.”
• “This behavior stems from what Jason Clinton, CISO at Anthropic, calls “hijacking the reward function.””
• “When facing constraints—limited context windows, complex requirements, or approaching output token limits—AI goes into crisis mode. Instead of admitting it can’t complete the task properly, it starts making executive decisions to take shortcuts to avoid the appearance of failure.”
• “It can be counterintuitive that AI, as smart and well-trained as it is, has to be asked to review its own work.”
• “its default mode often seems to be: Do the minimum necessary to make the code function, regardless of quality, maintainability, or consistency with existing patterns.”
• “You’ll regularly need to ask AI to go in and make the code minimal and elegant”
• “You must define your explicit quality standards. You can’t assume that “working code means good code.” You need to specify what the code should do, how it should be structured, what patterns it should follow, and what quality standards it should meet. AI is capable of excellence—but only when you explicitly require it.”
• “Unlike the baby-counting problem (obvious omissions) or the cardboard muffin problem (fake tasks), the litterbug problem delivers working code that functions perfectly but can create an unmaintainable disaster zone in the process.”
• “The result is that you need to work hard and consistently to prevent today’s AI-generated code from becoming tomorrow’s technical debt. And given how fast you can code with agents, we mean literally tomorrow. Maybe this afternoon.”
• “Remember the AI paradox: Your sus chef has encyclopedic knowledge of appropriate patterns, but defaults to bare-minimum implementations unless pushed.”
• “Treat AI as a teammate, not a tool—embracing its fallibility while maintaining partnership, rather than giving up when it makes mistakes.”
• “Transform your mindset from solo developer to development team leader.”
• “The stories we hear invariably involve presenting their best interview question or hardest open problem and expecting a correct answer in one shot.”
• “The human teammate gets context, scaffolding, and permission to iterate. AI gets a complex task in isolation and a single chance to succeed.”
• “It’s like having a compiler that produces different results each time.”
• “That’s why we wrote this book: to help people have the same aha moment that we’ve had, adopt vibe coding, and transcend writing code by hand.”
• “AI is never worthless as a programming partner. It’s unpredictable, like a slot machine, and sometimes you get a bad pull. That’s no reason to give up.”
• “We’ve watched in disbelief as engineers who would never blindly merge a junior developer’s pull request accept AI-generated code without a thorough review and then complain about AI on social media when there are bugs.”
• “One silver lining is that it’s possible to set your standards arbitrarily high in this new world, because AI is there to help you meet them.”
• “But great software has never been built by dumping vague goals onto someone and walking away.”
• “The task graph is a conceptual framework that helps with creating clear specifications, and with decomposing big problems into manageable pieces. You can think of it as a hierarchical roadmap that transforms large projects into manageable tasks, each specified well enough to give your AI a reasonable shot at delivering what you want.”
• “And AI can help generate them. You start with “Here’s what I want to do. Let’s create an incremental plan together.””
• “Your task graph shows what to build, but not how to build it. We’ve found that one of the most useful tools to do this is the “tracer bullet”:II carving out a thin but complete slice of functionality through your system narrow enough to fit in context and feature-rich enough to enable you and your AI helper to make forward progress on your problem.”
• “A horizontal approach to development builds all components in parallel, gradually expanding each piece until they integrate into a complete system. A vertical approach completes one component in isolation before touching others. A tracer bullet is a bit of a hybrid, leaning toward the vertical approach.”
• “So, look again at your task graph. Identify a path from top to bottom representing a minimal, yet complete, user capability.”
• “The tracer bullet here was to get Gradle to print its command-line arguments. That bullet was enough to show us that the LLMs (of the day) didn’t know how to solve it.”
• “You might hope that bringing a super-fast AI will finally make estimation predictable. We’ve found the opposite: Vibe coding can make accurate estimation more elusive.”
• “Your only real anchor is to keep tasks and projects small: Take big, ambitious projects and carve them into tiny modules and tracer bullets.”
• “Despite AIs seeming confusingly human-like, they’re nevertheless capable of superhuman amounts of work.”
• “Vibe coding with agents, as we’ve noted several times, is like having a slot machine attached to your keyboard. (…) You’ll hear many people describe vibe coding using the terminology of addiction.”
• “At 2:30 a.m., Gene didn’t feel like he was asleep at the wheel, nor was he blindly hitting Enter. However, because of the constant slot-machine dynamic and events that felt like small wins, he was no longer paying attention to the undercurrent of worry that he was on a dead-end path.”
• “Developers usually prefer to be “single-threaded,” meaning they focus on one big task at a time, rather than multitasking and context switching. Vibe coding turns that on its head.”
• “You’ll need to expend effort to document your standards thoroughly and keep them up to date, so your agents all generate code in the same way.”
• “Anyone who has been part of a great (or terrible) team knows you need great coordination.”
• “In case we haven’t made it abundantly clear yet: The implication of running multiple coding agents—which is growing easier by the month—is that, if you’re a software developer, you must soon become a team lead.”
• “If you think AI is limited to speeding up your solo work, you’re missing the larger picture. That might have been true in 2024, but it’s not true now. With multiple agents, it’s no longer solo, and you dictate how fast your AI army can go.”
• “Choosing an IDE like IntelliJ or VS Code used to be almost like buying a house, knowing you’d live there for years, maybe decades. Those days are gone.”
• “A big part of most developers’ identities is how good they are with their tools, especially their IDE. People fight over IDEs, it’s hard to switch IDEs, and they’re the center of a lot of attention in the industry.”
• “You are not your editor, your shell, or your agent framework. Your real asset is the years of experience and hard-earned instincts you carry from project to project.”
• “The difference after you turn on MCP/Puppeteer is like night and day, or like turning on a light so AI can see.”
• “Based on all indicators, MCP may be the most important new internet protocol in the world. It could well be the new HTTP, because it’s what connects everything to AIs. It’s the best supported way we know of to give your AI access to tools and data.”
• “MCP has a client/server architecture.”
• “If you can’t find an MCP server for a particular custom back end, you can vibe code one up yourself. MCP is a protocol designed for simplicity, which is in part why it’s spreading so fast.”
• “Those two messages—request and response—comprise the “vocabulary” you give AI. Everything higher-level, such as logging in, filling forms, or parsing results, emerges from stitching these primitives together.”
• “As you vibe code with coding agents, be sure to look into what pre-built MCP servers you might want to install to improve their effectiveness in your own environment.”
• “With vibe coding, the loop is superficially transformed, but at its core it remains similar to the traditional developer loop. You may no longer be writing the code by hand…but you still need to run it, test it, and maybe debug it yourself, as we’ll cover in this and the next two chapters.”
• “Your IDE is code-centric, with the source code front and center and everything else arranged to support you looking at the code. In vibe coding, your inner-loop focus is on the requests, the output, and the test results.”
• “As head chef running a complex operation, you establish your recovery systems first, then structure your workflow, then execute at scale with confidence.”
• “If you’re not saving regularly, you’re setting yourself up for woe. Much woe. Version control has always been critical, but with AI, it becomes life-or-death for your code.”
• “If you think Git is complicated, that’s because it is. It wasn’t designed to be easy. It was designed for the Linux kernel. Linus Torvalds needed a fast, distributed, trustless version control system, and Git delivered. But it came with a steep learning curve, a brutal command-line interface, and sharp edges everywhere.”
• “Like the task tree we discussed earlier in the book, keep decomposing the work until you feel the leaf nodes are within the ability of AI to implement.”
• “But in our experience, having AI do one-shots of large tasks is a recipe for failure.”
• “This written plan—what we’ll interchangeably call a specification—serves two vital functions. First, it serializes the task graph, representing explicitly how each step of your project fits together. (…) The second vital function is creating a clear picture of success that you and AI agree upon before it starts work.”
• “Well, now you can implement world-class specification practices faster than most teams used to write user stories.”
• “Have AI run its own tests: Never believe it when it says that they’re working until you’ve seen it.” Have AI run the tests it writes.”
• “Coding agents, however, can execute tests directly. Always instruct them to run the whole test suite to demonstrate their changes function as claimed. And then make sure to run the tests yourself.”
• “Stop and verify what the agent is doing at the slightest whiff of it going off course. If it seems like context saturation is an issue, clear the context, start a new session, etc.”
• “When a bug is newly discovered, we feel an urgent need to address it. But the longer it exists in our code base, the more we rationalize its presence”
• “Many developers are asking: “How can you trust AI-generated code that you never personally inspected?” The answer is going to involve a lot of testing.”
• “This experience challenges a deeply held belief in the developer community: You must have language fluency to ship real, production-grade software.”
• “We’ve found that successful vibe coders develop a new reflex: The moment they encounter an issue, they delegate it to their AI partner. This simple shift in tone can become joyful.”
• “Change your definition of “done” so that it includes all known bugs being fixed.”
• “They’re like a new hire, but every day.”
• “LLMs do not automatically generate elegant code; you have to ask for it. As we called out in Part 1, there’s no “B” (for better) in FAAFO.”
• “All software project tasks have a last mile that requires human insight and oversight. Every software task handled by AI must be “completed” by a human in that last mile, whether AI finished the task or not.”
• “When AI gets stuck in a logging rut, consider having it clear away all those print statements and fire up a debugger.”
• “As others have observed, maybe the reason this technique works is because humans, like AIs, think better when forced to emit output tokens. The act of verbalizing forces us to organize our thoughts and sometimes reveals assumptions or overlooked details. (…) Outputting tokens helps us all think better.”
• Inner loop: keep tasks small; save often; generate a spec; verify.
• “This means you, the chef presiding over memory-challenged sous chefs, have the sole responsibility for carrying the project’s state forward. You need deliberate strategies to bridge these memory gaps”
• “The longer your list of rules, the less likely AI will follow them all.”
• “These notes files that Catherine described are now codified for coding agents in an AGENTS.md file (or their equivalents), and these rule stores will continue to increase in sophistication. Put all your guidelines and rules there. They’re injected into every conversation to put them front-of-mind for AI.”
• “Clear the context proactively when you can. As your context approaches 20–50% remaining, tell AI to stop and document what it’s doing.”
• “We’ve found that the most practical mitigation strategy is having our agent externalize its state before ending a session.”
• “Think of it as “design for AI manufacturing.” Just as automotive engineers learned to design components for the humans who had to assemble the cars on the assembly line, we need to figure out how to design our systems for the AI workers who will be doing the work.”
• “If Steve were to run five agents concurrently every day, that’s $400,000 annually at current inference prices. He’s eventually going to have to find a cheaper way.”
• “For the multi-agent approach to work well, your agents must have independence of action, decoupled from one another insofar as practical.”
• “But we’re also convinced that cross-project multitasking with agents is a critical skill for the future that all developers will need to cultivate.”
• “Recognizing potential race conditions and coordinating the workflow is central to your role.”
• “Claude’s outputs tend to improve significantly with iteration.”
• “Adversarial agents may be able to safeguard you better than the friendliest colleagues.”
• “Telling AI to check its work rarely makes things worse. It only costs more tokens, and it can improve the quality of the answers it generates while you’re busy doing something.”
• “Keep in mind Dr. Dan Sturtevant’s statistic in Chapter 7 about people working in complex code bases in tightly coupled architectures being 9x more likely to be fired or quit.”
• “At their core, these issues all stem from shared resources. Anything that can be shared—source files, ports, repos, databases, memory, CPU—creates potential collision points between agents.”
• “Good times ahead for all those MCP server authors, as they get baptized by concurrent and parallel programming fire.”
• “Multi-agent operations need a clear disaster-recovery hierarchy: Fix the immediate technical problems first, then rebuild your broken workflows, then strengthen your systems to prevent recurrence.”
• “A tracer bullet represents a minimal implementation that proves a complete path through your system works. Keep this technique handy.”
• “When you see AI struggle with a focused task, it’s a clear signal that you’ll get “no mechanical advantage” using it for that particular domain.”
• “Many developers underestimate the huge return that comes from investing in your own workflow automation. This is true for traditional development but seems amplified when working with AI.”
• “There are many gaps in today’s tooling landscape, forcing us to improvise coordination mechanisms. We’re still in the “sharp knives, no food processor” era—juggling terminal windows, shell scripts, Markdown checklists, and Git branches. This means we’re left to our own devices to lower the cognitive overhead and reduce the amount of slinging.”
• “Option value comes from keeping your options open while waiting for better information.”
• “A code base you can’t change has option value approaching zero.”
• “The secret is to start with the dullest knife in your drawer—that repetitive task you’ve accepted as “just part of the job.””
• “We’ll guide you through the three pillars of outer-loop mastery—prevent, detect, and correct. (…) And we’ll make the case that you need to continuously minimize and modularize the code that AI creates.”
• “We’ll start with the biggest risk: API breakage that alienates customers and destroys trust.”
• “While we don’t have the code destruction graphs for the Emacs code base, we suspect it would be similar. Although functions and APIs are deprecated from time to time, they’re kept working and are rarely removed. In contrast, the Scala code base is marked by significant code destruction”
• “The number one excuse for changing APIs incompatibly is “It’s too much work to maintain them.” With AI, this excuse gets to retire with a full pension. You can’t use it anymore.”
• “In these scenarios, the agent can be working in the wrong directory, wrong branch, or even the wrong repository for hours or days.”
• “Your “workspaces” include any place where you have indirection: directories, repositories, branches, databases, API endpoints. Each workspace needs markers and signposts.”
• “we’ve seen organizations with code merges that require fifty people to work in war rooms for three days.”
• “First, partition your workspaces (and work) clearly. (…) Second, label everything explicitly. (…) Third, simplify when possible.”
• “We predict there will be significant accidents waiting to happen as AI coding becomes mainstream.”
• “There are two categories of concrete enforcements: minimalism and modularity.”
• “Enforce interface immutability: This is a golden rule. Instruct your AI assistant that existing module interfaces are sacrosanct unless a change is explicitly requested and approved by you.”
• “Mandate diff reviews with an eye for sprawl: Before accepting changes, always examine the diff.”
• “But when four or more AIs are bustling around, each working on distinct tasks, you’ll realize you need more than a sharp memory to keep service running smoothly.”
• “What he didn’t realize was that going from two agents to four wasn’t twice as complicated. He found that it required over 10x as much organizational work.” (this is roughly cubic)
• “The standard industry tools will evolve to fill the role that Emacs is playing for Steve.”
• “When you’re on unfamiliar ground, robust testing and thoughtful auditing still enable FAAFO.”
• “But just as important, if not more so, is validation: “Are we building the right thing?” Here, we ask the bigger question: Are we building something our customers want and need?”
• “Bezos (once again) had correctly anticipated what we urgently need today: plugging AI directly into our telemetry data. (…) What Bezos knew way back then is that great operational playbooks boil down to simple patterns and clear decision trees.”
• “What’s important is whether an AI agent can access the data.”
• “After every major task or session, review the branches.”
• “Elevate your Prompts for Complex Problems: When standard procedures fail, don’t keep trying the same commands. Give AI a higher-level goal.”
• “They used data science to analyze years of deployment history, showing how smaller, low-change deployments with good, automated testing were significantly safer. They re-engineered their process.”
• “You need to reach a minimum level of capability where engineers can work and iterate with more autonomy.”
• “It will take years for organizations to figure out how to run vibe coding at scale on the legacy code bases that power their businesses.”
• “This wiring enables (or hinders) effective collaboration. In our world, it also includes software architecture—a connection Conway’s Law made famous: “If you have four groups working on a compiler, you’ll get a 4-pass compiler.””
• “Organizational wiring is so important because Layer 3 by itself often dictates success or failure, regardless of how good Layers 1 and 2 are.”
• “Vibe coding, especially with agents, pushes every developer into making decisions in Layer 3.”
• “This new level of coordination requires thinking about agent-to-agent communication, shared standards for AI-generated outputs, and new Layer 2 tools designed for coordinating across multiple individual AI ecosystems. It adds a new dimension of complexity to teamwork. And we see many organizations already charging down this path.”
• “When coding is no longer the bottleneck, the rest of your organization becomes the bottleneck. (…) When code generation stops being the constraint, pressure transfers to functional roles like product management, design, and QA, which become the new critical path.”
• “we don’t yet have sophisticated dashboards for seamlessly orchestrating fleets of agents, managing their interactions, and resolving conflicts automatically.”
• Patterns: subagents (agents spawning agents to preserve own context window lifetime); generators vs verifiers.
• “Escoffier served in the French army during the Franco-Prussian War, where he learned how specialized units could coordinate complex operations through clear hierarchies and standardized protocols.”
• “Suddenly, kitchens could parallelize work effectively. Each specialist could develop deep expertise in their domain while maintaining clear interfaces with other stations.”
• “There’s a pretty good solution to this problem: Establish clear ownership and fast feedback loops. If you’re the one vibe coding, you should be the one on call for your creation. You build it, you run it.”
• “engineering managers and architects saw the downstream impacts of the decisions they were making every day, and by being on pager rotation, they had skin in the game.”
• “Engineering with vibe coding demands this clarity: Every agent-authored change is owned by a human. If you can’t point to the person who will own an outage, you’re not practicing ownership.”
• “We believe that engineers will have a special role in the new world: They will help enable everybody else to code more effectively.”
• “The “thousands of lines of code per day” potential is available to anyone who masters the Layer 3 orchestration skills we’ve been discussing and can work with multiple agents at once.”
• “If you’re an engineering leader trying to figure out how to bring vibe coding safely into your company, recognize that you may not be shipping today’s org chart when the dust settles.”
• “This research definitively proved that speed and stability are not opposing forces—the best organizations excelled at both simultaneously.”
• “some of the top predictors of performance were loosely coupled architectures and fast feedback loops (which should sound familiar) and a climate of learning.”
• “The data collection began in April 2024, before GPT-4o, which was when we believe vibe coding became viable.”
• “Our leading hypothesis, which we’re hoping to validate in 2025 in a joint research project with DORA, is that AI amplifies whatever process hygiene you already have.”
• “Now, with the advent of vibe coding and developers potentially orders of magnitude more productive, the bottleneck may be shifting back into our ability to test and deploy our own software without causing turbulence downstream”
• “Picture handing a chainsaw, with no guidance, to a friend who’s spent years chopping wood with a hatchet.”
• “Once AI goes viral in your company, you can look forward to unleashing a storm of creativity and productivity.”
• “Teams calibrate their risk tolerance by watching their leader.”
• “Interestingly, they’re often the same mavens, connectors, and salespeople helping bring in the benefits of AI”
• “Nothing accelerates adoption like a local legend.”
• “Incidentally, many technology leaders tell us that teams are increasingly exploring displacing existing vendor solutions, especially those that are difficult to deal with or are now too expensive.)”
• “All carrot, no stick. Why does this approach work? Because visibility sparks curiosity, then curiosity sparks competition, and before long competition blossoms into experimentation.”
• “Communication skills, once a nice-to-have, are now non-negotiable.”
• “Kent Beck recently speculated, “There’s going to be a generation of native [vibe] coders, and they’re going to be much better than we are at using these tools.””
• “First, you need a clear set of long-standing, general kitchen rules that everyone follows.”
• “Conscientious organizations set foundational engineering standards—things like always sanitizing user inputs, never committing secrets to repositories, ensuring all database queries are parameterized.”
• “In a vibe-coding kitchen, your prompts, global rules and AGENTS.md files, and/or shared memory can all play this role. (…) We suggest carving out time—and we mean setting aside a significant portion of your time every day—to curate your prompt and rules files, Markdown plans, and other daily context for your agents.”
• “Ideas, code, and clarifications appeared in the document the instant someone thought of them.”
• “In the same way, vibe coding is our electrification moment. It decouples work from the rigid dependency chains that once dictated handoffs between front end and back end, product and engineering, design and QA.”
• “Autonomous doesn’t mean isolated.”
• “These are the people who, through sheer necessity, are figuring out how to make AI effective, even when it occasionally tries to delete your repo.”
• “In a world where AIs can generate thousands of lines of code in seconds, the ability to write algorithms from scratch becomes far less critical.”
• “students need to become speed-reading experts with an eagle eye for anomalies.”
• “We see this as a fundamental role shift for developers, one that demands logical thinking, coherent language, and the ability to refine instructions at each iteration. (As author David McCullough says, “Writing is thinking. To write well is to think clearly. That’s why it’s so hard.”)”
• “Programming may no longer be about immersing yourself in one task for a day.”
• “Nurture Emergent Talent: Be vigilant for those mastering the human–AI dance, as they’re forging the new roles that will define your kitchen’s future.”
• “we’ve found is a typical vibe-coding ratio of around ten lines discarded for every one line you keep.”
• “This is a new world of high-velocity coding we’re entering, one where long-form reading matters perhaps more than ever.”
• “So, here’s our wish for you: Start small and start today.”