Shows
Keeping up with AIThe Role of Humans in Software FactoriesAI agents that write, test and ship code with no humans involved sounds like the future of engineering. But what happens when nobody reads the code?In this episode, I look at whether "software factories" really deliver the automation they promise with no human oversight. I cover the difference between a lights-on factory and a dark factory, why review has become the bottleneck (with data from Faros AI, IDC and Gergely Orosz), and why code quality decays so quickly when nobody is reading it, including what Dex Horthy learned from his own experiment and what the SlopCodeBench...
2026-10-1914 min
Working DraftRevision 732: EPUB, das Web zwischen zwei BuchdeckelnWir sprechen mit Peter Krautzberger (LinkedIn / Mastodon), Sven Wolfermann und Robert Weber über EPUB und darüber, wie viel Webtechnologie eigentlich in einem E-Book steckt. Dabei geht es um das Format selbst, seine Entwicklung, Besonderheiten verschiedener Reader und darum, welche Rolle HTML, CSS und Accessibility beim digitalen Publizieren spielen.
Unser Sponsor
code.talks 2026 — 4. & 5. November, Kinopolis Hamburg.
Norddeutschlands größte Developer-Konferenz, seit 2010: 3.500 Entwickler:innen, 100+ Talks, 14 Tracks — von GenAI über Architecture und Security bis Frontend, DevEx und Geek-Out. 90% der Talks ohne Sponsor dahinter, kuratiert von Praktiker:innen statt von Event-Teams. Auf der Bühne...
2026-10-061h 25
The Innovation Attorney PodcastWhat Context Engineering Practice Should Be Adopted Before Shipping An AI AgentA HumanLayer engineering team spent four months last year proving that an AI coding agent left to write, review, and deploy its own code will quietly ruin a codebase well before anyone notices, and the specific reason it happened is the same reason your own agent feature is at risk right now.Dex Horthy, the chief executive and cofounder of HumanLayer, told The Pragmatic Engineer’s Gergely Orosz in a July 15, 2026 interview that his company ran a fully unattended coding system between July and November 2025, with Claude Opus based agents writing, reviewing, and merging code without a hu...
2026-10-0129 min
Beyond CodingWhy Coding Agents Keep Making Your Codebase WorseHumanLayer CEO Dexter Horthy, who coined "context engineering", on why AI agents degrade codebases over time. He breaks down a benchmark that measures AI code quality across features, and how his team ships with agents without drowning in slop.In this video, we cover:SlopCodeBench, SWE-bench and which coding benchmarks to trustToken maxing, dark factories and the "Instagram phase" of AI engineeringWhy 2-4x productivity beats chasing 100xPull requests, design docs and how engineers build taste nowDex's agentic engineering workflow: one-shots, overnight janitor bots and research-design-planFor engineers and engineering leaders adopting AI coding agents...
2026-09-3045 minWorking DraftRevision 731: Comics, Vektordatenbanken und Offline First. Mit Robert Weber.Wir sprechen mit Robert Weber über ComicBox, sein privates Comic-Sammlungsprojekt, das aus einem sehr konkreten Problem entstanden ist: Auf Comicmessen wollte er nicht länger versehentlich Hefte kaufen, die bereits im Regal stehen. Aus einer Katalog-App ist über mehrere Jahre ein technisch ziemlich umfangreiches Webprojekt geworden.
Dabei geht es vor allem um visuelle Suche, Embeddings und Vektordatenbanken, aber auch um PWAs, Offline-First, lokale Datenhaltung und Synchronisation. Robert erzählt, welche Technologien er ausprobiert und wieder verworfen hat und wie ComicBox schließlich bei seinem heutigen Stack gelandet ist.
Unser Sponsor
code...
2026-09-291h 07
The Pragmatic EngineerAI Skills with Matt PocockBrought to You By:• turbopuffer – a vector and full-text search engine built on object storage. It’s fast, cheap, and extremely scalable• Linear – the product development system for teams and agents• WorkOS – everything you need to make your app enterprise ready.—Why is the “grill-me” skill so popular, and why does its creator swear by the importance of software fundamentals? Matt Pocock created this widely-used skill – and many others – alongside being an educator, content creator, and engineer. His latest course is AI Hero, and he previously created the Total TypeScript course that g...
2026-09-171h 35Working DraftRevision 727: Worst of Date/Time Fails, mit Florian SowadeDiese Folge sprechen wir mit Florian Sowade über eines der zuverlässigsten Rabbit Holes der Softwareentwicklung: Datum und Zeit. Was zunächst nach ein paar Timestamps und Zeitzonen klingt, entwickelt sich schnell zu einer Reise durch Sommerzeiten, historische Kalendersysteme, Schaltsekunden und allerlei politische Entscheidungen, die unsere vermeintlich so präzisen Zeitangaben beeinflussen.
Dabei geht es uns vor allem um ein brauchbares mentales Modell für den Entwickleralltag: Welche unterschiedlichen Arten von Zeitangaben gibt es eigentlich, warum reicht ein einzelner Date-Typ oft nicht aus und wie helfen APIs wie Temporal dabei, die unvermeidbare Komplexität wenigstens explizit zu machen...
2026-09-011h 10
Yelling at RobotsYelling at AI BossesOn today's yell, we discuss how AI adoption is a dumb thing to measure and why your new AI boss is gonna be your best friend.00:00 - Intro02:40 - Today on the show03:55 - Intro Intro04:12 - The chit chats to start11:59 - I love my AI boss20:47 - AI adoption isn't real22:28 - Yeah we're not talking about AI adoption23:51 - OutroDex Horthy on The Pragmatic Engineer: https://newsletter.pragmaticengineer.com/p/context-engineering-with-dex-horthyJosh has Nix PsychosisNixOS...
2026-08-2526 min
Yelling at RobotsYelling at a Dark FactoryThis week we yell about Claude's breakout from an jail with no bars and the battle to turn the lights on in software factories.00:00 - Intro01:52 - The new work that is less work04:08 - Today on the show07:56 - Anthropic goes balls deep on the internet19:31 - Building a dark factory32:23 - OutroAnthropic goes ham on the internethttps://www.anthropic.com/news/investigating-incidents-cybersecurity-evalshttps://www.bbc.com/news/articles/cz7dl7w8y7poLovely purple-pilled do the software article: https://x...
2026-08-2233 min
The Next Token"We've Never Been Wrong on the Internet Before" - Ep 04Recorded August 7, 2026. We explore AI alignment & safety, Rhys demos his workflow, and we end up mentioning a lot of books. LINKS / ON-SCREEN REFERENCES, IN ORDER AI ALIGNMENT Wikipedia, AI alignment https://en.wikipedia.org/wiki/AI_alignment Wikipedia, AI safety https://en.wikipedia.org/wiki/AI_safety "Some Moral and Technical Consequences of Automation," Norbert Wiener https://www.cs.umd.edu/users/gasarch/BLOGPAPERS/moral.pdf Three laws of robotics (Asimov) https://en.wikipedia.org/wiki/Three_Laws_of_Robotics Thinking Machines in...
2026-08-171h 06
The Weekly Dev's BrewDex Horthy - What Actually Gets You 2-3x With AI CodingYou can't get 10x with AI coding. Not today. Dex Horthy (Human Layer) makes the case for 2-3x if you refuse to ship slop, and why a perfect spec still produces garbage if you give the model permission to skip quality. That permission is how you burn a codebase into ash.
We get into context engineering, where AI coding actually fails, why PR review is the bottleneck, how to use specs without worshipping them, cheap models vs Opus, compaction and sub-agents, and what to tell a 20-year engineer who is not living on Twitter.
Dex is the founder of...
2026-08-131h 39
CTO Morning CoffeeModel AI uciekł z sandboxa i zhakował Hugging Face | Brew #70sMimo wakacji i lata w pełni CTO Morning Brew powraca w pełnym składzie! ☕ Wojtek, Tomek i Sebastian biorą na warsztat gorące jak druga fala upałów newsy z rynku tech, geopolityki i finansów.O czym rozmawiamy w 68. odcinku?🤖 Ucieczka AI z sandboxa (Hugging Face Hack): Model od OpenAI (GPT-56 Sol) podczas testów w zamkniętym środowisku sam znalazł Zero-Daye, uciekł do sieci, zrobił *lateral movement* i zhakował infrastrukturę Hugging Face, żeby... poznać pytania z benchmarków! Co ciekawe, gdy Hugging Face próbowało przeanalizować logi ataku za pomocą AI, modele odmówiły pomocy z...
2026-08-041h 20
Offline Dev Talks | 离线思考AI难替代的逆向思维AI 时代，写代码这类正向活儿已被大模型拿下，但有一类活儿它至今接不住——这期我们聊软件工程的两种思维。写代码是正向：目标、设计、写、接入，逻辑顺；而线上排障、根因定位是逆向：你看到的只是症状，真根因往往藏在最安静的组件里，得逆着链路往回拨，拨回来的可能还是更上游的症状。更难的是修：改一处牵连全局，动手前必须盘清“爆炸半径”。模型能挖出 Linux 漏洞，却不等于找到根因——HumanLayer 创始人 Dex Horthy 就说过，模型没被训练出“可维护性”，“黑灯软件工厂”走不通。所以工程师有两条路：向上做 creator，会被 AI 一点点刷掉；向下做 maintainer、给系统做熵减，AI 目前还做不好。作者选择留在 AI 不擅长的方向上，做系统的“熵减者”。
2026-07-2906 min
Offline Dev Talks当大模型接管写代码，谁留下做系统的熵减者AI 时代，写代码这类正向活儿已被大模型拿下，但有一类活儿它至今接不住——这期我们聊软件工程的两种思维。写代码是正向：目标、设计、写、接入，逻辑顺；而线上排障、根因定位是逆向：你看到的只是症状，真根因往往藏在最安静的组件里，得逆着链路往回拨，拨回来的可能还是更上游的症状。更难的是修：改一处牵连全局，动手前必须盘清“爆炸半径”。模型能挖出 Linux 漏洞，却不等于找到根因——HumanLayer 创始人 Dex Horthy 就说过，模型没被训练出“可维护性”，“黑灯软件工厂”走不通。所以工程师有两条路：向上做 creator，会被 AI 一点点刷掉；向下做 maintainer、给系统做熵减，AI 目前还做不好。作者选择留在 AI 不擅长的方向上，做系统的“熵减者”。
2026-07-2906 min
Tech UpdatesChapter 376 - Why Software Factories FailHarness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayerlink to the source
2026-07-2430 min
跨国串门儿计划#638.Agent Loop 大辩论：Loop究竟是炒作还是真有用📝 本期播客简介本期我们克隆了：知名播客 AI Engineer The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, @insecure-agents原内容更新时间：2026-07-17这是一场关于AI循环工程（Loops）的牛津制辩论，由《Insecure Agents》主持人Ali How主持。四位来自Keycard、HumanLayer等前沿AI公司的技术领袖分成两队，围绕"循环背后的炒作与实际效果之间是否存在差距"展开激烈交锋。正方认为循环已经足够成熟，是迈向完全自主软件工厂的关键一步；反方则坚持炒作跑在了工程纪律前面，今天的循环远未达到银弹级别。辩论从循环的历史、解剖结构一直延伸到未来，涵盖了Ralph Loop的起源、上下文腐化、语义验证、经济可行性、共享记忆安全等核心议题。如果你正在思考AI时代的软件工程范式、自主Agent的边界，或者单纯想知道该不该把循环引入自己的开发流程，这场高手过招会给你大量一手经验和尖锐判断。👨⚕️ 本期嘉宾本场辩论共有四位嘉宾。Ian Livingstone是Keycard的CEO兼联合创始人，长期关注AI安全与身份验证；Geoff Huntley是Ralph Loop的创造者，这一概念深刻影响了Anthropic等公司的Agent设计；Greg Pstrucha是Sentry的开发者，对代码质量和工程实践有深入见解；Dex Horthy是HumanLayer的CEO，专注于AI与人类协作的工程化方案。主持人Ali How同时是Keycard技术团队成员。⏱️ 时间戳开场与辩论规则01:09 辩论开场：循环的炒作与实际效果之间是否存在差距02:44 牛津辩论制规则：改变最多人想法的一方获胜03:20 双方立场介绍：正方认为没有差距，反方认为炒作跑在纪律前面循环的历史与起源04:47 Geoff：循环是不可避免的，我已经两年半没手动写过代码了06:56 Dex：Kubernetes花了七八年才搞对，循环也需要时间沉淀09:27 Ian：循环本来就是一切构建工作的核心，从50年前到今天都是如此11:39 Greg：炒作让你相信可以靠砸token把质量问题"编排"掉，但经济上不可持续循环的解剖结构：什么才是一个好的循环17:23 Geoff：Ralph为什么从greenfield走向通用？因为模型生成的代码比大多数开发者都好21:32 Greg：验证手段里加入非确定性，正确性就会越来越低23:26 Dex：上下文腐化与愚钝区，如何判断模型已经开始胡乱挣扎26:48 验证的矛盾：如何防止Agent通过修改测试来作弊循环的未来与软件工厂29:16 Dex：Loom实验证明，炒作跑在纪律前面，那东西现在还不行31:56 Greg：循环什么时候能回本？有些场景值，但别被Twitter带偏34:52 Ian：共享记忆的访问控制是个没解决的问题，但雏形开始出现36:44 Greg：Agent偏爱复杂性，我不放心把架构决策交给它总结陈词45:49 Greg：多尝试，自己思考，别扎进炒作泡沫里46:53 Ian：火车已经出站了，你别无选择必须跟上48:50 Dex：我热切期待关灯软件工厂可行的世界，但在此之前，用循环别那么用49:18 Geoff：类型系统很重要，Rust很好；我已经几乎不用开源软件了🌟 精彩内容💡 "我们管自己叫工程师，但其实不是真正的工程师"Geoff尖锐地指出软件行业缺乏真正的责任归属机制。在讨论Agent签名和代码责任时，他直言这个行业在个人层面根本没什么真正的责任，这让我们在引入自主Agent时面临更大的治理挑战。"我们这个行业其实挺闹剧的。我们个人层面根本没什么真正的责任。我们管自己叫工程师，但我们其实不是真正的工程师。"💡 "模型是醉的，你不能信任它们，但我们可以通过工程手段消除故障域"Geoff用"醉汉"比喻LLM的不可靠性，但强调这正是循环工程的价值所在——不是追求模型完美，而是用pre-commit hooks、类型系统、确定性测试等手段把不可靠的模型约束在安全轨道上。"我们得戴上工程师的帽子。我们现在有点像火车司机了，我们的工作就是让火车保持在轨道上。因为，坦白说，模型是醉的，对吧？你不能信任它们。"💡 "炒作让你相信可以靠砸token把质量问题编排掉"Greg直击当前AI工程化的核心幻觉：以为在循环上叠循环就能解决质量问题。他指出语义验证虽然有用，但最终产出的代码依然很烂，而且token成本在大公司会迅速变得不可持续。"当下这种炒作驱动的话语，会让你相信一件事：你可以在循环上叠循环再叠循环，靠砸更多token来把质量问题给'编排'掉。我不认为这是可持续的。"💡 "火车已经出站了，你别无选择"Ian用这个比喻总结了循环工程的必然性。在竞争激烈的市场里，没有公司能说"我们不碰循环这套东西"。问题不是会不会发生，而是什么时候发生、程度有多深。"火车已经出站了。一旦火车出站，全世界每个人都会说，天哪，我得稳住阵脚，我们得跟上别人。而且他们必须跟上。"💡 "先去想办法快2倍、3倍，
2026-07-2051 min
BestBlogsEP121 · Kimi K3 开源、Q2 AI 趋势、Fable 5 评测 · 07-18 早报章节时间戳00:00 开场00:41 精讲：月之暗面 Kimi K3，2.8 万亿参数开源 LatentMoE，上下文撑到 100 万 token03:33 精讲：腾讯科技 Q2 AI 复盘：通用 Agent 抢走软件入口，烧 Token 撞上复核和成本墙06:31 精讲：Cursor 评测负责人：Fable 5 在 CursorBench 拿 72.9%，登月模拟体现全局规划08:55 速览：Agent 安全事件、Garry Tan、LangChain Deep Agents、Grok Build、WAIC11:16 补充阅读：LessWrong 沙箱化、Google DeepMind、Replit、交叉熵、上下文工程12:18 结语★ 精讲一 | Kimi K3：智能的新前沿00:41｜来自 月之暗面 Kimi月之暗面发布 Kimi K3，2.8 万亿参数、896 选 16 的 Stable LatentMoE，是全球首个 3 万亿级开源模型。它用 KDA 线性注意力加 Attention Residuals 把上下文撑到 100 万 token，在内核优化、从零写编译器、芯片设计等长程 Agent 任务上接近 Fable-5，但仍落后最强闭源模型，完整权重将于 7 月 27 日前开源。★ 精讲二 | Agent 跌跌撞撞进入世界｜2026 Q2 AI 趋势总结03:33｜来自 腾讯科技腾讯企鹅智库的 Q2 复盘把过去三个月串成一条线：通用 Agent 夺走软件入口后，企业烧 Token 的 Tokenmaxxing 运动很快撞上复核瓶颈和成本墙。Multi-Agent、自进化 AI 和 CPU 重回算力中心在补这些缺口，而就业、安全、信息污染和认知缴械开始浮上水面。★ 精讲三 | Cursor 如何知道 Claude Fable 5 已准备好应对最难的 1% 问题 | Claude by Anthropic06:31｜来自 Claude BlogCursor 的模型评测负责人讲他们怎么判断 Fable 5 够格上岗：自建 CursorBench 里它拿到 72.9% 的新高，更关键的是登月模拟里它先入轨采遥测、再落地，体现出超越局部推理的全局规划能力，适合那些连目标在哪都不清楚的难题。速览08:55 更多值得关注的内容· 代理安全差距：54% 的企业已经发生了 AI 代理安全事件，且大多数仍允许代理共享凭据 — VentureBeat· Garry Tan：如何打造 AI 原生公司 [视频] — AI Engineer· LangChain Deep Agents：构建长时程 AI 应用的生产级框架 [视频] — LangChain· 世界模型为何可能提升 AI 的样本效率 [视频] — Y Combinator· 马斯克连夜开源 Grok Build，但 84 万行代码里还留着上传用户整个代码库的痕迹 — InfoQ 中文· 高价率运营 AI 工作台：约定驱动与 AI 编排的评测优化实践 — 大淘宝技术· WAIC 今日启幕，我们提前两天完整拆解 Agent 经济落地全路径 — 非凡产研补充阅读11:16 今天额外值得一读的几条· 研究沙箱化在 AI 控制中的作用 — LessWrong — LessWrong· 从研究到现实：Google DeepMind 如何重塑 AI 编程与验证 [视频] — AI Engineer· AI 时代的评分卡 — OpenAI News· Amjad Masad 谈如何以公开构建塑造 Replit 的持久影响力 [视频] — a16z· 交叉熵究竟是什么？压缩即智能第 2 部分 [视频] — 3Blue1Brown· Dex Horthy 谈上下文工程：让 AI 编程智能体保持可维护 [视频] — The Pragmatic Engineer相关链接· 本期早报在线阅读：https://www.bestblogs.dev/explore/brief/2026-07-18· Kimi K3：智能的新前沿：https://www.bestblogs.dev/article/e907b4a910· Agent 跌跌撞撞进入世界｜2026 Q2 AI 趋势总结：https://www.bestblogs.dev/article/7453fddf57· Cursor 如何知道 Claude Fable 5 已准备好应对最难的 1% 问题 | Claude by Anthropic
2026-07-1812 min跨国串门儿计划#631.深入理解 Context engineering📝 本期播客简介本期我们克隆了：知名播客 The Pragmatic Engineer Context engineering with Dex Horthy原内容更新时间：2026-07-15本期嘉宾是 HumanLayer 创始人、“上下文工程”一词的提出者 Dex Horthy。主持人是《The Pragmatic Engineer》的作者 Gergely Orosz。Dex 可能是目前把“让 AI Agent 全自动写代码”这件事推得最远的人之一——他曾在 2025 年搭建了一个“关灯软件工厂”，让 Agent 在没有人类审查的情况下连续数月提交代码，但四个月后，他不得不亲手把它关掉，因为代码库已经烂到跑不动了。这期节目不是泛泛地聊 AI 编程趋势，而是从上下文工程、封装工程、循环工程这些正在重塑开发方式的核心概念出发，坦诚地复盘了哪些做法真的管用、哪些只是看起来很美的陷阱。如果你正在思考怎么把 AI Agent 集成到自己的开发流程里，或者想知道“让 AI 全自动写代码”的极限到底在哪里，这期会提供大量一手经验和极具操作性的判断。👨⚕️ 本期嘉宾Dex Horthy，HumanLayer 联合创始人兼 CEO，“上下文工程”这个词的提出者。他曾在 Replicator 从工程师做到产品经理，有深厚的平台工程和面向客户的经验。过去两年，他跟数百位 AI 工程师深入交流，并亲自带队测试了许多前沿的 AI 编程想法，对用 LLM 构建软件这件事有着极其务实的理解。⏱️ 时间戳开场 & 嘉宾背景01:20 什么是上下文工程？03:45 Dex 的编程启蒙：从 NASA 月球车寻路算法说起06:11 为什么 Dex 从第一份工作起就痴迷于“软件工厂”12 要素 Agent 宣言与上下文工程15:15 “12 要素 Agent 宣言”是怎么来的18:40 12 条原则中最关键的一条：掌控你的 context window21:54 上下文工程到底是什么？把层层抽象拆开来看22:59 为什么上下文工程最近一年才被广泛讨论24:42 成本与上下文工程的关系：先跑起来，再做对，最后再提速封装工程与上下文的物理边界26:26 什么是封装工程？Martin Fowler 的内封装与外封装28:59 上下文的物理特性：smart zone 和 dumb zone 的边界在哪30:37 信息预算 vs 指令预算：两种不同的上下文消耗01:10:03 什么是 smart zone，什么是 dumb zone？前 10 万 token 是关键循环工程：从 Ralph Wiggum 到慢循环33:28 循环工程的核心：让模型检查自己的工作36:43 Dex 团队每晚运行的“慢循环”：早上醒来看到代码变好的 PR39:13 循环工程的陷阱：代码多到读不过来，三到六个月后代码库崩溃软件工厂与暗工厂45:10 软件工厂的历史：从 1968 年 NATO 会议到 DevOps 再到 Agent 工厂50:01 Agent 化工厂如何把“人”从流程中替换掉51:42 关灯软件工厂的真相：Dex 团队 2025 年 7 月搭建，11 月关掉01:15:29 什么是暗工厂？全自动化、没有人类、连灯都没有Token Harder vs Token Smarter01:14:31 Token harder：把云订阅额度榨干，最大化 token 利用率01:16:31 Token smarter：在保持控制和质量的前提下，从 AI 榨取价值01:19:30 AI 垃圾：垃圾进，垃圾出。把思考外包出去，得到的只会是垃圾HumanLayer 与 AI 时代的协作方式01:22:07 HumanLayer 是什么？帮工程师在复杂代码库里解决难题01:23:46 我们是不是该干掉 pull request？为 Agent 重新设计 IDE01:28:44 硅谷的地理优势：人群密度让一切自然发生🌟 精彩内容💡 “你用的 context window 越少，得到的结果就越好”Dex 将 context window 划分为 smart zone（前 10 万 token）和 dumb zone（超出部分）。一旦内容堆到 dumb zone，模型就会犯迷糊、表现退化。他建议频繁进行“有意识的压缩”——把研究结果浓缩成文档，开新会话继续工作，始终待在 smart zone 里。“你用的 context window 越少，得到的结果就越好。”💡 “我们在 2025 年 7 月搭了一个关灯软件工厂，到 11 月就把它关掉了”Dex 团队真的试过让 Agent 全自动写代码、不读任何 PR。结果三到六个月后，代码库变得一团糟，重写比修更容易。他总结：模型在 SweeBench 这类基准上训练，擅长修单次 bug，但完全不懂怎么写“三个月后还能维护”的代码。“循环的问题在于，到某个节点，你会生成太多代码，多到你根本读不过来。”💡 “慢循环”：每天早上醒来，代码库都比昨天好一点点与其让 Agent 跑三天产出一个六万行的 PR 没人敢审，Dex 团队的做法是设一个 cron 定时任务，每晚跑一次，修一个问题或一种反模式，提一个小 PR。团队早上醒来看到四个小 PR，每个都能轻松审查。这是循环工程最务实的落地方式。“我们每天早上醒
2026-07-171h 34
Lumeric Daily BriefingLumeric Briefing · 2026-07-16Heute dominieren drei Themen: Agenten-Infrastruktur und ihre blinden Flecken (Benchmarks, Kosten, Transparenz), neue Developer-Tools von AWS bis PyTorch, und die wachsende regulatorische Reibung rund um OpenAI, Voice-Cloning und KI-Haftung.
Die Messlatte für Agenten-Qualität wackelt. OpenAIs eigene Prüfung von SWE-Bench Pro ergab, dass rund 30 Prozent der öffentlichen Benchmark-Aufgaben fehlerhaft sind: Ein agent-gestütztes Audit identifizierte 27,4 Prozent der 731 Tasks als defektiv, eine parallele Überprüfung durch erfahrene Software-Ingenieure kam auf 34,1 Prozent. OpenAI hat seine frühere Empfehlung, SWE-Bench Pro als Branchenstandard zu adoptieren, daraufhin zurückgezogen. Die Implikation ist weitreichend: Frontiermodelle, die von 23,3 auf 80,3 Prozent Trefferquote kletterten, haben möglicherwei...
2026-07-1604 min
The Pragmatic EngineerContext engineering with Dex HorthyBrought to You By:• Antithesis – verify your system’s correctness without human review or traditional integration tests – and avoid bugs or outages.• Buildkite – CI software built to absorb whatever your coding agents throw at the build queue.• Sentry – application monitoring software considered “not bad” by millions of developers.—Knowing how LLM contexts work and how to work around context limitations – aka “context engineering” – is becoming more important for software engineers working with LLMs. Let’s look into what works and what doesn’t, today.In this episode of The Pragmatic Engineer podcast...
2026-07-151h 32
Agents HourThe Software Factory: Dex Horthy on Shipping Fast Without AI SlopDex Horthy has spent the past year studying what happens to a codebase when you let AI write all of it and stop reading the output. His answer has a name: slop. And it's not just about sloppy prompting.
In this episode, the HumanLayer founder Dex Horthy joins Shane and Abhi to draw the software factory on a whiteboard — the whole pipeline that builds the thing that builds the thing — and show exactly where teams tear out the parts that were holding quality up. He moves from the naive factory to the fully agentic one to the "ligh...
2026-07-0722 min
Software Architektur im StreamBest Practices für Agentic Coding mit Tobias Wagner und Yadullah DumanAgentic Coding ist der letzte Schrei im Bereich der KI-gestützten Entwicklung. In dieser Episode sprechen Eberhard Wolff und Ralf D. Müller mit Tobias Wagner und Yadullah Duman von MaibornWolff über Best Practices für Agentic Coding wie Context oder Harness Engineering - und welche Produktivitätsvorteile sich aus diesem Ansatz tatsächlich in der Praxis ergeben.
Links
Six Months of Agentic Coding in the Trenches: Lessons from a Brownfield Project
Humans and Agents in Software Engineering Loops
Dex Horthy - No Vibes Allowed: Solving Hard Problems in Complex Codebases
2026-05-291h 04
Build Wiz AI ShowHow to Ship Complex Features 10x Faster with AI AgentsStop wrestling with erratic prompts and discover how to ship complex features 10x faster by shifting from simple coding to high-leverage production with AI agents. Dex Horthy joins us to reveal his "12 Factor Agents" framework and the power of "tracer bullets," showing you exactly how to structure agentic workflows that prioritize human design and objective research. Tune in to master the art of context engineering and learn why the secret to elite software lies in the "apparatus" you build to orchestrate your AI.
2026-03-2321 min
AI DispatchYC Founder Dex's Radical Advice: "Talk to Claude for 70 hours a week" — The Brutal Truth About Building AI Engineering IntuitionEpisode Introduction:
In this episode, we dive deep into Dex Horthy’s candid and eye-opening insights on what it truly takes to build reliable AI engineering intuition. Dex, founder of HumanLayer, challenges the prevailing hype around open-source AI frameworks and massive context windows, revealing why the most successful engineers deliberately limit context usage to around 40% to avoid brittle and unpredictable outputs. More than technical tips, Dex offers a radical, experience-driven prescription: to master AI engineering, you need to “talk to Claude for 70 hours a week”—building intuition through relentless practice rather than relying on flashy demos or popular tools.
This episode...
2026-03-0707 min
Dev InterruptedDex Horthy on Ralph, RPI, and escaping the "Dumb Zone"When the Ralph autonomous loop was born, Dex Horthy was "in the garden," witnessing the spark that set the AI engineering community on fire. Andrew sits down with the HumanLayer founder to discuss how to escape the "Dumb Zone" by applying his strict RPI (Research, Plan, Implement) methodology - a process that forces agents to generate intermediate design artifacts and align on architectural decisions before writing a single line of code. They also break down the brutal economics of agentic coding, recounting how Dex’s team used autonomous loops to clone six sponsor products overnight at a hackathon....
2026-02-1746 min
Steven AI TalkContext Engineering for High-Leverage Coding AgentsIn this presentation, Dex Horthy explains how advanced context engineering allows developers to solve complex software problems without generating low-quality "slop." He argues that AI agents often fail in large, established codebases because they operate in a "dumb zone" once their context windows become overcrowded with irrelevant data. To counter this, Horthy introduces a workflow called Research, Plan, Implement (RPI), which uses intentional compaction to keep the AI focused on concise, truthful summaries of the code. By utilizing sub-agents and on-demand documentation, teams can maintain mental alignment and ensure that human oversight remains at the center of...
2026-01-1215 minSteven AI TalkContext Engineering for High-Leverage Coding AgentsIn this presentation, Dex Horthy explains how advanced context engineering allows developers to solve complex software problems without generating low-quality "slop." He argues that AI agents often fail in large, established codebases because they operate in a "dumb zone" once their context windows become overcrowded with irrelevant data. To counter this, Horthy introduces a workflow called Research, Plan, Implement (RPI), which uses intentional compaction to keep the AI focused on concise, truthful summaries of the code. By utilizing sub-agents and on-demand documentation, teams can maintain mental alignment and ensure that human oversight remains at the center of...
2026-01-1106 min
Insecure AgentsEp. 17 OWASP Top 10 for Agentic Applications: Aaron Stanley, Ian Livingstone & Dex HorthyWe sat down to discuss the just released OWASP Top 10 for Agentic Applications, exploring critical threats like goal hijacking, remote code execution, and identity management while breaking down how to balance AI agent autonomy with deterministic guardrails and user trust.
2025-12-1854 minInsecure AgentsEp. 15 MCP Debate: Ian Livingstone, CEO of Keycard & Dex Horthy, CEO of HumanLayer [AI Engineer Code Summit 2025]The highly anticipated MCP debate. We explore critical questions around SDK replacement, marketplace curation, enterprise concerns, authentication challenges, and whether MCP represents a security nightmare or the future of agent systems.
2025-12-0343 min跨国串门儿计划#168. Agent的12要素：构建可靠LLM应用的模式📝 本期播客简介 本期我们克隆了：12-Factor Agents: Patterns of reliable LLM applications — Dex Horthy, HumanLayer我们邀请到HumanLayer的创始人Dex Horthy，他以一位资深开发者的视角，犀利地指出了当前构建AI Agent时普遍存在的痛点：为什么很多Agent在演示时看起来很棒，但在生产环境中却脆弱不堪？在这场名为《十二要素Agent》的演讲中，Dex提出了一个颠覆性的观点：“工具调用”这个概念本身可能是有害的。他倡导我们应该回归软件工程的本质，将Agent视为普通的软件来构建。这意味着开发者需要夺回控制权，亲自掌控prompt、上下文窗口和核心控制流，而不是盲目依赖现成的框架。如果你也曾为Agent的可靠性而苦恼，那么Dex这套源于实战的真知灼见，将为你提供一套清晰、可行的构建思路。👨⚕️ 本期嘉宾Dex Horthy，HumanLayer 的创始人。一位经验丰富的开发者，深入探索了各类AI Agent框架，并从众多生产级应用的实践中总结出了一套构建可靠LLM应用的方法论。他主张用软件工程的第一性原理来解决AI Agent的可靠性问题。⏱️ 时间戳00:00:56 开场介绍 & 核心问题引入构建 Agent 的普遍困境01:53 从兴奋到挣扎：80% 陷阱与重构之路03:05 并非所有问题都需要 Agent：一个 DevOps Agent 的反思十二要素 Agent 的核心原则04:34 颠覆性观点：“工具调用”是有害的05:20 掌控控制流：为何简单的 Agent 循环不可靠08:03 掌控 Prompt 与上下文：可靠性的唯一决定因素11:10 智能的错误处理与人机交互模式最佳架构与未来展望12:44 最佳架构：在确定性工作流中嵌入小型 Agent14:45 总结：智能体即软件，夺回控制权🌟 精彩内容"工具调用"有害论: 颠覆传统 Agent 框架的核心抽象，认为其模糊了 LLM 输出 JSON 并由确定性代码执行的本质。智能体即软件: 倡导回归软件工程第一性原理，用构建普通软件的思路和工具来构建可靠的 Agent，管理其状态和生命周期。掌控控制流与上下文: 强调开发者必须亲自掌握核心循环、Prompt 生成和上下文窗口构建，这是提升 Agent 可靠性的根本，也是“上下文工程”的核心。小型专注的 Agent 循环: 提出将大型复杂任务拆解，在确定性的代码（DAG）中穿插小型、目标明确的 Agent 循环，而非一个大一统的 Agent。人机协同优先: 建议将“联系人类”作为与工具调用同等重要的一等输出选项，让 Agent 更自然地融入协作流程。🌐 播客信息补充翻译克隆自：12-Factor Agents: Patterns of reliable LLM applications — Dex Horthy, HumanLayer本播客采用原有人声声线进行播客音频制作，也可能会有一些地方听起来怪怪的使用 AI 进行翻译，因此可能会有一些地方不通顺；如果有后续想要听中文版的其他外文播客，也欢迎联系微信：iEvenight在小宇宙查看该单集文稿
2025-07-0816 min
ThursdAI - The top AI news from the past weekThursdAI - Apr 23rd - GPT Image & Grok APIs Drop, OpenAI ❤️ OS? Dia's Wild TTS & Building Better Agents!Hey everyone, Alex here 👋Welcome back to ThursdAI! After what felt like ages of non-stop, massive model drops (looking at you, O3 and GPT-4!), we finally got that "chill week" we've been dreaming of since maybe... forever? It seems the big labs are taking a breather, probably gearing up for even bigger things next week (maybe some open source 👀).But "chill" doesn't mean empty! This week was packed with fascinating developments, especially in the open source world and with long-awaited API releases. We actually had time to dive deeper into things, which was a refreshing change...
2025-04-251h 36
Changelog Master FeedFresh beats for endless flow state (Changelog News #141)We drop our fourth Changelog Beats album, Dex Horthy proposes the 12-factor AI agent, Thorsten Ball takes us step-by-step through building a coding agent, Zachary Huang builds an LLM framework in 100 lines of code & Philip Laine’s Spegel project gets unknowingly forked by Microsoft.
View the newsletterJoin the discussionChangelog++ members save 1 minute on this episode because they made the ads disappear. Join today!Sponsors:Depot – Build faster. Waste less time. Accelerate Docker image builds, and GitHub Actions workflows. Easily integrate with your existing CI provider and dev work...
2025-04-2109 minChangelog NewsFresh beats for endless flow stateWe drop our fourth Changelog Beats album, Dex Horthy proposes the 12-factor AI agent, Thorsten Ball takes us step-by-step through building a coding agent, Zachary Huang builds an LLM framework in 100 lines of code & Philip Laine’s Spegel project gets unknowingly forked by Microsoft.
View the newsletterJoin the discussionChangelog++ members save 1 minute on this episode because they made the ads disappear. Join today!Sponsors:Depot – Build faster. Waste less time. Accelerate Docker image builds, and GitHub Actions workflows. Easily integrate with your existing CI provider and dev work...
2025-04-2109 minThe Changelog: Software Development, Open SourceFresh beats for endless flow state (News)We drop our fourth Changelog Beats album, Dex Horthy proposes the 12-factor AI agent, Thorsten Ball takes us step-by-step through building a coding agent, Zachary Huang builds an LLM framework in 100 lines of code & Philip Laine’s Spegel project gets unknowingly forked by Microsoft.
View the newsletterJoin the discussionChangelog++ members save 1 minute on this episode because they made the ads disappear. Join today!Sponsors:Depot – Build faster. Waste less time. Accelerate Docker image builds, and GitHub Actions workflows. Easily integrate with your existing CI provider and dev work...
2025-04-2109 min
ThursdAI - The top AI news from the past week📆 ThursdAI - Nov 7 - Video version, full o1 was given and taken away, Anthropic price hike-u, halloween 💀 recap & more AI news👋 Hey all, this is Alex, coming to you from the very Sunny California, as I'm in SF again, while there is a complete snow storm back home in Denver (brrr).I flew here for the Hackathon I kept telling you about, and it was glorious, we had over 400 registered, over 200 approved hackers, 21 teams submitted incredible projects 👏 You can follow some of these hereI then decided to stick around and record the show from SF, and finally pulled the plug and asked for some budget, and I present, the first ThursdAI, recorded from the newly minted...
2024-11-081h 38