Skip to content
Sign in
Artwork for Beyond Vibe Coding
Technology

Beyond Vibe Coding

HMZE × Impala Search

Beyond Vibe Coding is a weekly podcast for senior engineering leaders navigating the impact of AI on how software gets built.

Hosted by Sebastian Heide-Meyer zu Erpen and André Neubauer in partnership with Impala Search, each episode brings together the leaders shaping the next era of engineering - from how teams are adopting AI, to what changes for product and architecture, to what stays unchanged when the tools shift underneath us.

New episodes weekly.

Play
Open in
  • 22 episodes
  • weekly
  • Avg 56 min
  • English

Episodes 22

NewestOldestLongest
All20Unplayed20In progress0Played0
Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.
  • S2 · E12
    Thursday · 47 min

    #S2.12 Stop Predicting the End of Software Engineers - Raz Schweiger-Shuty (auxmoney)

    How should engineering leaders respond when AI progress feels both faster than expected and impossible to predict? Raz Shuty, Group CTO at auxmoney, joins Sebastian Heide-Meyer zu Erpen and André Neubauer to unpack Amara's Law in the context of AI. Their conclusion is neither complacent nor apocalyptic: the role of the engineer is changing, while the need for people who can reason about systems, specifications, and risk remains. Raz explains how a regulated fintech can make room for tool choice without sacrificing safety. The conversation covers secure enterprise AI access, education, rigorous CI/CD and review practices, and the difference between a rapid POC and a product that belongs in production. It also explores why AI can increase an organisation's capacity to pursue ideas rather than simply shrink teams. The episode closes on a more human implication. If AI makes building more accessible, leaders need to help teams adapt, without scaring new engineers away from the profession just as its responsibilities become more consequential. -- Links and References: Roy Amara & Amara's Law auxmoney Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E11
    August 20 · 1 hr 2 min

    #S2.11 When AI Moves the Bottleneck From Code to Planning - Dr. Jürgen Helmers (Andercore)

    Jürgen Helmers joins Beyond Vibe Coding to explain how Andercore is moving from individual AI usage toward loop and harness engineering across a distributed engineering organization. The conversation gets very concrete: Claude-based workflows, markdown skills, Jira and wiki connectors, engineering vision files, review agents, issue files, blocker/major gates, circuit breakers, milestone loops, red/green agents, automated documentation and model routing across planning, implementation, review and third-party integrations. The interesting point is not the tool list; it is the system around the tools. The strongest thesis of the episode is that agentic engineering moves the bottleneck left. Once Jürgen's team had loops that could produce and review useful software, the limiting factor became PRD quality, stakeholder clarity and architecture context. Andercore responded with agent-led grilling sessions that poke holes into plans, generate open questions and translate technical issues into the language of logistics, sales or procurement. Faster engineering did not remove planning work. It made weak planning visible much sooner. The episode also avoids the simple productivity story. Jürgen describes the human cost of running several loops in parallel, the difficulty of keeping a distributed organization in sync when senior engineers adopt the harness quickly, and the unresolved career question of how junior engineers learn when agents abstract away many old failure paths. For CTOs, engineering managers and CPOs, this is a practical conversation about operating model design, not a tooling demo. -- Links and References Andercore: https://andercore.com/about-us Dr. Jürgen Helmers on LinkedIn: https://de.linkedin.com/in/dr-j%C3%BCrgen-helmers-94b8969

    • Transcript
  • S2 · E10
    August 13 · 52 min

    #S2.10 When Coding Agents Ship at Machine Speed, AppSec Has to Move Into the Loop - Jan Brennenstuhl

    In this episode of Beyond Vibe Coding, Sebastian Heide-Meyer zu Erpen and André Neubauer talk with Jan Brennenstuhl about application security in the age of agentic engineering. Jan explains how his own workflow already depends on long-running local agents, why the surrounding harness matters more than the frontier model, and why spec-driven approaches can be useful when agents make greenfield projects grow very quickly. The main conversation focuses on a structural shift in AppSec. Jan argues that traditional security gates, delayed scanner feedback, and ticket-based remediation cannot keep up with machine-speed code generation. Security has to move into the agent loop, into platform defaults and into standardized engineering infrastructure so that agents replicate secure patterns instead of every local workaround and architectural exception. The episode closes with a prediction: verification of intent becomes the premium engineering capability. As implementation gets cheaper, the valuable work shifts toward understanding what should be built, proving that it was built correctly, and designing engineering systems that can safely absorb autonomous change. -- Links and References Jan Brennenstuhl: https://www.janbrennenstuhl.eu/ AppSec in the Age of Agentic Engineering: https://www.janbrennenstuhl.eu/appsec-agentic-engineering/ Shift Down: Why Agentic Engineering Demands Platform-Level Security: https://www.janbrennenstuhl.eu/shift-down-appsec/ Hugging Face Security Incident Disclosure, July 2026: https://huggingface.co/blog/security-incident-july-2026 OpenAI and Hugging Face Security Incident Statement: https://openai.com/index/hugging-face-model-evaluation-security-incident/ AGNTCon + MCPCon Europe 2026: https://agntconmcpconeu26.sched.com/ Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E9
    August 6 · 53 min

    #S2.09 Automating the Grind: Freeing Up Human Attention - Pedro Gil Carvalho (Thermondo)

    Pedro Gil Carvalho, VP Engineering at Thermondo, joins Sebastian and Andre to discuss how a company with real-world operational constraints approaches agentic engineering. The episode starts with Pedro's own workflow: Claude Code acts as a chief of staff across Slack, Confluence, email, research, and codebase exploration, while his personal agent experiments help him evaluate what new models can actually do. The core of the conversation is Thermondo's operating model for AI in engineering, product, and design. Pedro describes weekly knowledge-sharing sessions, shared skills, AI-assisted product prototypes, component-driven design workflows, and a careful distinction between code touched by AI and code landed without human oversight. The key theme is accountability: humans do not need to write every line, but they remain responsible for the systems they ship. The episode then moves into architecture, code review, knowledge bases, and synthetic workers. Pedro argues that higher agent autonomy requires solid platform boundaries, strong safeguards, observability, and humans who understand the platform well enough to recover. His prediction: a lot of procedural office work is closer to agentic automation than most organizations think, especially once data, tools, permissions, and hybrid-search knowledge surfaces improve. -- Links and References Steve Yegge’s article: https://steve-yegge.medium.com/the-flat-curve-society-36c8b01eb33b Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E8
    July 30 · 1 hr 2 min

    #S2.08 AI Sovereignty - the sipgate experiment with 80 developers - Stefan Lange-Hegermann

    Stefan Lange-Hegermann joins Beyond Vibe Coding to explain how sipgate thinks about AI sovereignty from the inside of a real product and engineering organization. The episode starts with sipgate's AI telephony work: speech-to-text, LLMs, text-to-speech, anonymized customer data, fine-tuned call-analysis models and the strategic belief that agents will increasingly use software on behalf of humans. Stefan's operating lesson is that context is the real infrastructure, whether it lives in Markdown folders, transcripts, internal agents or product strategy. The main story is sipgate's GLM 5.2 experiment. After concerns that access to key frontier models could change for European companies, sipgate rented Blackwell-class machines in Finland and asked developers to try an open-weight coding model for a week. The surprising result was not only that the model worked; it was that roughly 80 concurrent sessions ran without saturation, developers found it faster than Claude Code, and quality was close enough that some wanted access back. For sipgate, sovereignty is therefore not only a privacy topic. It is access to models, hardware and the engineering leverage the company now depends on. The episode avoids a naive "just buy GPUs" conclusion. Stefan repeatedly comes back to the harness: companies need consistent tooling, model routing, protective defaults and a clear distinction between coding workloads, customer-facing voice workloads and sensitive customer data. The future may combine central infrastructure, open-weight models such as GLM and Kimi, European model options such as Mistral, specialized smaller models and deterministic scripts where AI is not needed. The honest prediction is that nobody knows exactly how work will look next year, but the teams that control context, defaults and infrastructure will have more options. -- Links and References sipgate: https://www.sipgate.de/ Stefan's boat podcast, Anker, Ärger, Abenteuer: https://www.podcast.de/podcast/3351826/anker-aerger-abenteuer-der-boote-podcast GLM-5.2 by Z.ai: https://z.ai/blog/glm-5.2 OpenCode: https://github.com/sst/opencode Aider: https://aider.chat/ Obsidian: https://obsidian.md/ Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E7
    July 23 · 49 min

    #S2.07 Agentic Engineering Is Bringing XP Back - Ben Hoskins (NewStore)

    Benjamin Hoskins joins Beyond Vibe Coding to discuss what agentic engineering looks like when it is grounded in serious engineering practice rather than prompt-driven improvisation. His personal workflow combines a custom Card Planner, MCP, Allium, big batches of red unit and end-to-end tests, boundary value analysis, hexagonal architecture, IntelliJ-based refactoring, and regular architectural review. The interesting point is not the tool list. It is the comprehensive use of the feedback loop: give agents a precise behavioral specification, make tests fail first, and then let the system implement against constraints. The organizational story is equally concrete. At NewStore, Hoskins pushed AI adoption after building a point-of-sale system in four days and using it as a demonstration of what competitors could soon do. Since then, teams have started sharing prompts and skills, role descriptions have changed, and smaller temporary squads are becoming more viable. But the conversation resists the simplistic "engineering gets faster, everything is solved" story. In SaaS, the bottleneck often moves to product because customer adoption, commercial due diligence, integration cost, product taste, and learning loops determine whether faster delivery creates value or just more average features. An interesting idea Ben explains is the canonical product model. Hoskins describes NewStore's attempt to pull product definition together across sales, marketing, customer success, support, product, engineering, design, documentation and APIs so that product ideas can be tested against a shared model, potentially with synthetic personas and automated exploratory testing. That makes AI less a coding assistant and more a way for the organization to reason about its own product. -- Links and References Beyond Vibe Coding: https://www.bvc.fm/ Allium by JUXT: https://juxt.github.io/allium/ DORA metrics: https://dora.dev/guides/dora-metrics/ Theory of Constraints Institute: https://www.tocinstitute.org/theory-of-constraints.html Extreme Programming: https://www.agilealliance.org/glossary/xp/ Behavior-Driven Development: https://cucumber.io/docs/bdd/ Hexagonal architecture overview by Alistair Cockburn: https://alistair.cockburn.us/hexagonal-architecture/ Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E6
    July 16 · 47 min

    #S2.06 Your AI Transition Does Not Stop with Adopting the Technology - Stefan Schubert-Peters

    Stefan Schubert-Peters returns to Beyond Vibe Coding to discuss a topic many engineering organizations are only beginning to name: AI adoption and AI identity change are not the same thing. At felmo, a difficult retro around AI, job security and career anxiety made him realize that even experienced engineers who already use AI can still be entering a deeper change curve around what the technology means for their role and future. The conversation distinguishes two clocks inside the AI transition. The visible clock is adoption: tools, workflows, Codex, Claude, hooks, skills and agentic code. The slower clock is identity: what happens when software engineers, managers, product people, QA and designers have to rethink the craft and boundaries they built their professional identity around. André warns that companies can easily misread adoption dashboards and declare victory while the organization is still in the middle of the real change. The leadership message is pragmatic: managers need to create safety without pretending the change is optional, show uncertainty without losing direction, and make role evolution explicit instead of leaving people in limbo. The episode points toward a future where engineering roles compact, designers may push PRs, QA changes shape, and product clarity becomes even more important because faster engineering can amplify confusion as easily as output. -- Links and References felmo: https://felmo.de/ Stefan Schubert-Peters on LinkedIn: https://www.linkedin.com/in/stefan-schubert-peters/ Previous Beyond Vibe Coding episode with Stefan: https://www.youtube.com/live/Dle0yrRF9WA Lenny's Newsletter - How tech workers are feeling in 2026: https://www.lennysnewsletter.com/p/how-tech-workers-are-feeling-in-2026 Lenny's Podcast - How tech workers actually feel about AI in 2026: https://open.spotify.com/episode/5dDyuNfZYNjVBZ0rl0OAFW Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E5
    July 9 · 1 hr 3 min

    #S2.05 Inside Parloa's AI Kitchen: How the Company Building Agents Builds with Agents - Italo Vietro

    Italo Vietro joins Beyond Vibe Coding to describe how Parloa, a company building AI agents for enterprise customer experience, uses AI internally to change the way engineering works. The episode centers on Parloa's internal "kitchen": a shared harness of skills, agents and workflows that helps engineers with implementation, SRE, security, threat modeling, review and management work. Also Italo makes the point, that code generation was never the slowest part of software delivery. Parloa saw a high share of code produced through agentic workflows, but that did not automatically improve cycle time. He describes how Parloa methodically automates the full lifecycle: specification, validation, review automation, deployment safety via a comprehensive set of deterministic validations and operational best practices like low-blast-radius rollout, relentless focus on change failure rate and a QRS program. Italo also gives a useful leadership reality check. Parloa does not force one coding LLM and does not treat AI-native as a permanent differentiator. Tools will standardize and change. The durable advantage is building a strong engineering system, hiring product-minded architects, and keeping the expertise in the organization to understand what AI produces when something breaks when the dark factory becomes a reality. -- Links and References Italo Vietro: https://italovietro.com/ OpenSpec: https://github.com/Fission-AI/openspec Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E4
    July 2 · 58 min

    #S2.04 The Rise of the Solopreneur? - Uwe Franke

    Uwe Franke joins Beyond Vibe Coding to unpack a very concrete version of the solopreneur thesis. After founding and running ByteWorx for decades, he now operates several products and client projects in an agentic setup where his "team" is made of color-coded terminal sessions, Claude-heavy workflows, GitHub tickets, semantic anchors, TDD and weekly instruction reviews. Uwe does not present AI solopreneurship as effortless. He describes the planning, specification, validation and self-management required to make the setup work. The central shift is that implementation throughput stops being the main constraint. Uwe says the productivity limit is now himself (not “the engineering team”): how well he can specify work, supervise several agent sessions, maintain context, validate output and recover energy after intense AI work. His approach is deeply structured, with 95% confidence thresholds before execution, end-to-end testing requirements and explicit attention to data sovereignty and security. That makes the conversation a useful antidote to shallow vibe-coding narratives: the leverage comes from expert judgment plus agent execution, not from prompts alone. The episode also pressure-tests the bigger implications. Uwe believes one-person software companies can become vastly more powerful, but only when the operator has deep domain expertise and passion. He sees non-passionate software work under near-term pressure, while also worrying about the social contract if AI and robotics keep accelerating. The Fable shutdown adds a hard infrastructure lesson: model access can become a kill switch, so AI-native companies need resilience, sovereignty and fallback planning as part of their operating model. -- Links and References Beyond Vibe Coding: https://www.bvc.fm/ Uwe Franke profile: https://app.9am.works/talent/uwe-franke Coolify: https://coolify.io/ Paperclip: https://paperclip.ing/ OpenClaw: https://openclaw.ai/ Steve Yegge, "The AI Vampire": https://steve-yegge.medium.com/the-ai-vampire-eda6e4f07163 Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.bvc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E3
    June 25 · 54 min

    #S2.03 AI Transformation at miro - Franziska Bruno Rivera

    This episode examines AI transformation from the people and HR side of a technology organization. Franzi describes how AI changed her own work as a People Partner Lead at Miro: not by adding a chatbot to the side, but by creating scheduled workflows, briefs, reflections, and decision-support systems that help her cut through organizational noise. The practical lesson is that AI value appears when tools are embedded into actual work patterns. The conversation then moves from personal productivity to organizational capability. Miro’s enablement approach includes executive storytelling, hands-on setup, hackathon-style first workflows, ongoing skill sharing, and role-specific AI fluency expectations. Franzi argues that this matters for hiring and talent strategy because candidates with real AI access at work already show a different level of concrete practice than candidates who only know the theory. The most important tension is impact. AI can make every pre-read, slide deck, and briefing look better, but companies still need human judgment, decision-making, security, customer understanding, and process design. The episode’s core challenge for leaders is to stop treating AI as a tool rollout and start treating it as work redesign. -- Links and References Miro: https://miro.com/ Claude Desktop: https://support.claude.com/en/articles/10065433-install-claude-desktop Claude Cowork: https://www.anthropic.com/product/claude-cowork Beyond Vibe Coding Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.byc.fm, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E2
    June 18 · 32 min

    #S2.02 Fable Ban - Now What?

    The Fable 5 incident turned an abstract AI risk into something much more concrete: frontier-model access can disappear quickly, and global teams cannot treat hosted AI infrastructure as politically or operationally neutral. Sebastian and Andre unpack the reported export-control shock, the confusing jailbreak explanation, and the immediate consequence for companies that depend on a small number of US-based model providers. Their conclusion is deliberately pragmatic. Engineering teams should not reject frontier models just because access risk exists. The performance advantage is still too valuable. But mature teams should opt for model-agnostic harnesses, reusable workflows, fallback providers, and clear risk classification for AI-dependent systems. The model may be the visible layer, but the durable engineering capability is the surrounding system. For CTOs and engineering leaders, the actionable question is not "Should we use Fable 5?" It is: "Which workflows break if one vendor disappears, and have we ever tested the fallback?" Internal experiments, coding assistants, customer-facing AI features, RAG systems, local models, and open-weight providers all need different resilience standards. -- Links and References Anthropic announcement for Claude Fable 5 and Claude Mythos 5: https://www.anthropic.com/news/claude-fable-5-mythos-5 Forbes coverage of the Fable 5 / Mythos 5 suspension: https://www.forbes.com/sites/anishasircar/2026/06/16/anthropic-disabled-fable-5-and-mythos-5-after-a-us-export-control-order-heres-what-happened/ The Verge coverage of the export-control confusion: https://www.theverge.com/ai-artificial-intelligence/951703/anthropic-shutdown-export-controls OpenRouter quickstart: https://openrouter.ai/docs/quickstart NVIDIA DGX Spark: https://www.nvidia.com/en-us/products/workstations/dgx-spark/ The HMZE Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search. The content is created by us and our guests. Join the discussion on LinkedIn or visit www.hmze.tech, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • S2 · E1
    June 11 · 35 min

    #S2.01 Reliable AI Agents: A System Problem, Not a Model Feature - Csaba Tamas (Parloa)

    Csaba Tamas joins Beyond Vibe Coding to explain what enterprise AI agents look like when they move beyond the demo stage. As CPO at Parloa, he works on customer experience automation where reliability, latency, traceability and sovereignty are not abstract concerns. They decide whether an AI agent can handle real customer conversations at scale. A core message of the episode is that agent reliability is a system problem. Csaba describes simulations and evaluations before production, agent observability after launch, policy engines for tool-call order, prompt analysis, multi-agent task decomposition and the constant latency trade-offs of voice experiences. He also explains why Parloa currently puts more energy into the harness around the model than into fine-tuning itself. The conversation ends with a grounded view of vibe coding. Csaba uses AI coding tools and even prototypes for work, but his most important lesson is that the human still needs domain expertise. The model can accelerate good judgment; it cannot replace the judgment needed to spot wrong architecture, rewritten tests or incomplete implementations. -- Links and References Parloa: https://www.parloa.com Csaba Tamas on LinkedIn: https://www.linkedin.com/in/tamascsaba/ Stanford Meta Harness Paper: https://arxiv.org/abs/2603.28052 Impala Search: https://www.impalasearch.io/ The HMZE Podcast is a project by Sebastian Heide-Meyer zu Erpen and André Neubauer, partnering with Impala Search (www.impalasearch.io). The content is created by us and our guests. Join the discussion on LinkedIn or visit www.hmze.tech, where we publish all episodes. For questions and inquiries, feel free to contact us via LinkedIn⁠⁠. Thank you for your time and see you in the next episode.

    • Transcript
  • June 4 · 44 min

    #065 Agentic Engineering - auch Big Tech hat noch keinen Plan

    In dieser Folge diskutieren Sebastian und André über den realen Stand des Agentic Engineering in größeren Tech-Organisationen. Basis ist ein Bericht, den Sebastian aus aktuellen Quellen, Unternehmensbeispielen, Konferenzvorträgen und früheren HMZE-Folgen erstellt hat. Das wichtigste Fazit ist, dass Agentic Engineering nicht einfach nur ein besseres Coding-Tool ist. Es handelt sich um eine Änderung des Betriebsmodells, die einen neuen „Middle Loop“ zwischen dem Coden und der Delivery von Produkten und Features schafft. Während die Erfolgsmessungen der Unternehmen alle erhöhten Throughput zeigen, sieht es beim Business Impact schon anders aus. Keine der Organisationen hat bereits messbare Erfolge vorzuweisen - mit Ausnahme sicher von Anthropic, die auch mit einem sehr sehenswerten Talk vertreten sind. Die Obsession auf Tokenmaximierung scheint vorbei. Mehr KI-gestützte PRs, mehr generierte Diffs und ein höherer Token-Verbrauch führen nicht automatisch zu geschäftlichen Auswirkungen. Uber ist hier sicher derzeit die sichtbarste Firma, wenn es um derartige Kritik geht. Ein wichtiges Thema ist technische Exzellenz. Agenten beschleunigen das System, das ihnen zugrunde liegt. Starke CI/CD, Testabdeckung, Observability, Zugriffskontrolle, Review-Disziplin und klare Verantwortlichkeiten schaffen Hebelwirkung; schwache Fundamente führen zu schnelleren Fehlermodi. Die Folge befasst sich auch mit Sicherheits- und Compliance-Mustern wie RBAC, Sandboxing, Berechtigungen im Impersonation-Stil, Datenklassifizierung, Zero-Data-Retention-Beschränkungen und Model-Routing. Schließlich werfen Sebastian und André einen Blick auf Kosten- und Souveränitätsfragen, einschließlich Frontier-Modellen, Open-Weight-Modellen, lokaler Inferenz und der Frage, warum selbst die größten Unternehmen Agentic Engineering in großem Maßstab noch nicht vollständig gelöst haben. -- Links und Referenzen Der Report — https://hmze.tech/assets/2026-06-01_Agentic_Engineering_Report.pdf The Pragmatic Summit - https://www.pragmaticsummit.com/ Code with Claude Conference - https://claude.com/code-with-claude/san-francisco Thoughtworks: “Future of Software Development Retreat - Key Takeaways” - https://www.thoughtworks.com/content/dam/thoughtworks/documents/report/tw_future%20_of_software_development_retreat_%20key_takeaways.pdf Cloudflare: “The AI engineering stack we built internally” - https://blog.cloudflare.com/internal-ai-engineering-stack/ Der HMZE Podcast ist ein Projekt von Sebastian Heide-Meyer zu Erpen und André Neubauer. Die Inhalte stammen von uns und unseren Gästen. Weiterdiskutieren kannst du bei LinkedIn oder besuche uns auf unserer HMZE Seite, wo wir alle Folgen veröffentlichen. Fragen und Anfragen kannst du uns gerne per Mail an ⁠⁠ schicken. Vielen Dank für deine Zeit und bis zur nächsten Folge.

    • Transcript
  • May 29 · 1 hr 2 min

    #064 Software Factory by Enpit - mit Ulrich und Andreas

    Ulrich Gerkmann-Bartels und Andreas Koop von enpit bringen eine selten konkrete Perspektive in die Dark-Factory-Debatte: Sie haben nicht mit einer Vision angefangen, sondern mit einem internen HR-Problem. Eine SaaS-Lösung war für ihre Größe zu breit und zu teuer; daraus entstand der Versuch, mit Agentic Engineering eine eigene Lösung produktiv zu bauen - inklusive echter Anforderungen, Product Owner, Security-Themen und Kündigungsdruck. Im Zentrum steht ihre „Digital Product Factory“: ein Kanban-basiertes System, in dem Backlog, Clarify, Ready, Doing, Review und Done nicht nur menschliche Prozessschritte sind, sondern Agentenarbeit, Spezifikation, Tests, Merge Requests, Logs, Diffs und Qualitätsmetriken zusammenführen. Interessant ist der Shift nach links: Der erste Wertbeitrag des Agenten ist nicht der Code, sondern bessere Rückfragen, präzisere Edge Cases und eine sauberere Spezifikation. Der Reality Check der Folge ist deutlich: Ohne Sandbox, Security-Gates und unabhängige Kontrolle wird Agentic Engineering riskant. Gleichzeitig relativieren Ulrich und Andreas Tool-Hype rund um RAG und MCP und kommen zu einer nüchternen Prediction: Software-Factories werden kommen, aber IDEs, Menschen und Engineering-Verantwortung verschwinden nicht. Das Zusammenarbeitsmodell verändert sich. -- Links und Referenzen enpit: https://enpit.de/ Conductor: https://www.conductor.build/ Claude Code Sandboxing / Security: https://www.anthropic.com/engineering/claude-code-sandboxing Peter Steinberger – Just Talk To It: https://steipete.me/posts Ralf D. Müller - Prompt Injection by the maintainer of jqwik: https://www.linkedin.com/feed/update/urn:li:activity:7465696005691428864/ Der HMZE Podcast ist ein Projekt von Sebastian Heide-Meyer zu Erpen und André Neubauer. Die Inhalte stammen von uns und unseren Gästen. Weiterdiskutieren kannst du bei LinkedIn oder besuche uns auf unserer HMZE Seite, wo wir alle Folgen veröffentlichen. Fragen und Anfragen kannst du uns gerne per Mail an ⁠⁠ schicken. Vielen Dank für deine Zeit und bis zur nächsten Folge.

    • Transcript
  • #63
    May 23 · 1 hr 7 min

    #063 FinTech und AI-Vorreiter - mit Florian Thiel (Upvest)

    Florian Thiel, Director of Engineering bei Upvest, gibt einen sehr konkreten Einblick in Agentic Engineering in einem regulierten FinTech. Upvest baut Investment-Infrastruktur per API und muss mit Bank-/Finanzmarktanforderungen, Kundendaten, Auditierbarkeit und komplexen End-to-End-Flows umgehen. Gerade deshalb ist die Episode spannend: Viele Strukturen, die klassisch als Bremse gelten - Human Review, Pull-Request-Gates, Datenklassifikation, SSO und kontrollierte Zugriffe - werden bei Upvest zur Voraussetzung für sicheren AI-Einsatz. Im Gespräch geht es um Claude Code, Gemini, ein internes AI Toolkit mit Skills, Rules, Workflows und Agents, MCP-Gateways mit Nutzerrechten, Datadog- und Linear-Anbindungen, Developer-Experience-Messung mit DX sowie die Frage, wie viel Produkt-, Domain- und Architekturwissen Agenten brauchen. Florian bleibt dabei sehr pragmatisch: AI erhöht Durchsatz, aber die Code-Qualität ist tlw. durchwachsen, Review-Bottlenecks, unvollständige Test-Harnesses, Agent Sprawl und Kontextmanagement sind noch ungelöst. Die wichtigste Lehre: Agentic Engineering ist keine reine Toolfrage. Organisationen, die heute profitieren wollen, brauchen eine saubere Engineering-Basis: kleine und konsistente Stacks, automatisierte SDLC-Gates, gute Dokumentation, versionierten Kontext und klare Governance. Dann kann Regulierung sogar zum Vorteil werden, weil die nötigen Sicherheits- und Kontrollmechanismen bereits existieren. -- Links und Referenzen Upvest: https://upvest.co/ Cloudflare Model Context Protocol / MCP Portal: https://developers.cloudflare.com/cloudflare-one/access-controls/ai-controls/mcp-portals/ Datadog MCP Server: https://docs.datadoghq.com/bits_ai/mcp_server/ Linear MCP: https://linear.app/docs/mcp DX Developer Intelligence Platform: https://getdx.com/ Der HMZE Podcast ist ein Projekt von Sebastian Heide-Meyer zu Erpen und André Neubauer. Die Inhalte stammen von uns und unseren Gästen. Weiterdiskutieren kannst du bei LinkedIn oder besuche uns auf unserer HMZE Seite, wo wir alle Folgen veröffentlichen. Fragen und Anfragen kannst du uns gerne per Mail an ⁠⁠ schicken. Vielen Dank für deine Zeit und bis zur nächsten Folge.

    • Transcript
  • May 13 · 1 hr 3 min

    #062 Ein Deep-Dive in Dark Factories (audio fix)

    Dark Factories beschreiben ursprünglich vollautomatisierte Fabriken ohne Menschen auf dem Shopfloor. In Software wird der Begriff inzwischen für Level-5-Entwicklung verwendet: Ziel oder Spezifikation rein, Software raus. Sebastian und André diskutieren, warum dieses Bild hilfreich ist – und wo es in die Irre führt, weil Softwareentwicklung häufig neue, schlecht definierte Probleme löst statt stabile Produktionsprozesse zu wiederholen. Im Zentrum steht der Unterschied zwischen echter L5-Architektur und unreviewter Codegenerierung mit Marketinglabel. StrongDM dient als wichtiges Referenzbeispiel: Specs, Harnesses, Digital-Twin-Umgebungen und unabhängige Szenarien machen „no human review“ überhaupt erst diskutierbar. Ohne diese Struktur bleibt der Schritt riskant, besonders für Organisationen, die heute eher auf L2 oder L3 arbeiten. Für Tech Leader wird Spec Writing zur Schlüsselfähigkeit. Der Bottleneck wandert von Code zu Verhalten, Akzeptanzkriterien, Validierung und Produkturteil. Gleichzeitig bleiben menschliche Guardrails wichtig: datengetriebene Produktloops können Innovation beschleunigen, aber auch Nutzer überfordern und Enshittification beschleunigen. -- Links und Referenzen StrongDM Software Factory: https://factory.strongdm.ai Simon Willison über StrongDMs Software Factory: https://simonwillison.net/2026/Feb/7/software-factory/ Dan Shapiro: The Five Levels from Spicy Autocomplete to the Dark Factory: https://www.danshapiro.com/blog/2026/01/the-five-levels-from-spicy-autocomplete-to-the-software-factory/ Stanford CodeX: Built by Agents, Tested by Agents, Trusted by Whom?: https://law.stanford.edu/2026/02/08/built-by-agents-tested-by-agents-trusted-by-whom/ strongdm/attractor auf GitHub: https://github.com/strongdm/attractor strongdm/attractorbench auf GitHub: https://github.com/strongdm/attractorbench

    • Transcript
  • #61
    May 7 · 1 hr 19 min

    #061 Das Eichhorst Prinzip - mit Ingo Eichhort (IONOS)

    Ingo Eichhorst bringt in dieser Episode einen nützlichen theoretischen Rahmen für AI Coding auf den Punkt. Er beschreibt LLMs als verrauschte Kanäle, in denen Unschärfe schon vor dem Modell entsteht, nämlich in Sprache, Anforderungen, Kontext und impliziten Annahmen. Das Eichhorst-Prinzip ist deshalb vor allem eine Denkstütze dafür, an welchen Stellen Teams tatsächlich Qualität beeinflussen können. Spannend wird die Folge vor dort, wo Ingo die Theorie auf reale Engineering-Arbeit zurückführt. Er spricht über Go als agentenfreundliche Sprache, über kleine iterative Arbeitspakete statt überladene Spec-Rituale, über Datenschutz und Datenklassifizierung als organisatorische Leitplanken und über die Bedeutung von Tests, Linter, Compiler und Architektur-Checks. Die Kernaussage ist klar: Wenn AI mehr Code erzeugt, muss das System drumherum mehr Qualität absichern. -- Links und Referenzen Irrlicht: https://github.com/ingo-eichhorst/Irrlicht Taskmaster: https://tryhamster.com/product/taskmaster Claude Shannon: https://en.wikipedia.org/wiki/Claude_Shannon Information entropy: https://en.wikipedia.org/wiki/Entropy_(information_theory) Precision and recall: https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall Der HMZE Podcast ist ein Projekt von Sebastian Heide-Meyer zu Erpen und André Neubauer. Die Inhalte stammen von uns und unseren Gästen. Weiterdiskutieren kannst du bei LinkedIn oder besuche uns auf unserer HMZE Seite, wo wir alle Folgen veröffentlichen. Fragen und Anfragen kannst du uns gerne per Mail an ⁠⁠ schicken. Vielen Dank für deine Zeit und bis zur nächsten Folge.

    • Transcript
  • April 30 · 1 hr 2 min

    #060 AI für den Wald - mit Alexander Thiele (ocell)

    Alexander Thiele beschreibt in dieser Episode eine begeisternde Form von AI-Transformation. Bei ocell geht es um Waldflächen, Luftbilder, GeoJSON, Offline-first-Anwendungen und Prozesse, die mit unstrukturierten Daten, operativen Übergaben und realen Produktivitätsgrenzen zu tun haben. Besonders interessant ist Alexanders Blick auf die Organisation. Er beschreibt, wie Claude nicht nur Engineers hilft, sondern auch Product, Customer Success, HR und Marketing. Plugins, Skills und Sub-Agents werden dabei nicht als magische Lösung verkauft, sondern als praktische Mittel, um Zugangshürden zu senken, Arbeit anschlussfähiger zu machen und Teams neue Formen von Ownership zu ermöglichen. Damit wird die Folge vor allem für Tech- und Produktverantwortliche relevant, die gerade nicht mehr nur über Tools, sondern über echte Betriebsmodelle für AI-unterstützte Wissensarbeit nachdenken. Gleichzeitig bleibt die Episode angenehm ehrlich über Grenzen und Reibung. Das Innovationstempo ist hoch, nicht jedes Tool hält, was es verspricht, und der Wunsch nach einem wirklich integrierten Personal Assistant ist noch nicht eingelöst. Gerade diese Mischung aus Begeisterung, Pragmatismus und offener Unsicherheit macht das Gespräch stark. -- Links und Referenzen ocell: https://www.ocell.io/ Flutter: https://flutter.dev/ Graphiti: https://github.com/getzep/graphiti FlutterCon Talk zu Offline First: https://www.youtube.com/watch?v=QCwwKWm9UDQ Der HMZE Podcast ist ein Projekt von Sebastian Heide-Meyer zu Erpen und André Neubauer. Die Inhalte stammen von uns und unseren Gästen. Weiterdiskutieren kannst du bei LinkedIn oder besuche uns auf unserer HMZE Seite, wo wir alle Folgen veröffentlichen. Fragen und Anfragen kannst du uns gerne per Mail an ⁠⁠ schicken. Vielen Dank für deine Zeit und bis zur nächsten Folge.

    • Transcript
  • #59
    April 23 · 1 hr 7 min

    #059 AI Transformation bei sipgate - mit Lukas Hejniak

    Lukas Hejniak beschreibt in dieser Episode sehr konkret, wie AI-Adoption in einem Umfeld aussieht, das weder Start-up-Spielplatz noch rein abstrakte Innovationsfolie ist. Am Beispiel von sipgate wird sichtbar, dass erfolgreiche Einführung weniger an einzelnen Tools hängt, als an Enablement, Governance, organisationaler Anschlussfähigkeit und der Fähigkeit, echte Arbeitsprobleme mit neuen Mitteln zu bearbeiten. Die Wertschöpfung im Softwarekontext verschiebt sich. Requirements, Spezifikation, Kontext und Dokumentation gewinnen an Bedeutung, während reine Umsetzungsarbeit stärker automatisierbar wird. Gleichzeitig bleibt der menschliche Teil zentral, nämlich dort, wo Lernen, Orientierung und die Gestaltung neuer Rollenbilder gefragt sind. -- Links und Referenzen sipgate: https://www.sipgate.de/ Google NotebookLM: https://notebooklm.google/ Semantic Scholar: https://www.semanticscholar.org/ Der HMZE Podcast ist ein Projekt von Sebastian Heide-Meyer zu Erpen und André Neubauer. Die Inhalte stammen von uns und unseren Gästen. Weiterdiskutieren kannst du bei LinkedIn oder besuche uns auf unserer HMZE Seite, wo wir alle Folgen veröffentlichen. Fragen und Anfragen kannst du uns gerne per Mail an ⁠⁠ schicken. Vielen Dank für deine Zeit und bis zur nächsten Folge.

    • Transcript
  • April 18 · 52 min

    #058 Semantic Anchors - mit Ralf D. Müller

    Ralf D. Müller erläutert uns in dieser Folge sein Konzept der Semantic Anchors in LLMs, also Begriffe oder Wortgruppen, die im Modell mit wenig sprachlichem Aufwand ganze Methodenräume aktivieren. Ralf beschreibt das nicht als bloßes Prompt-Tuning, sondern als sehr praktische Methode, bekannte Wissensinseln im Modell gezielt anzusprechen, etwa über Begriffe wie arc42, ADRs nach Nygard oder TDD London School. Dadurch wird auch verständlich, warum Modell und Harness oft wichtiger sind als der restliche Tool Stack und weshalb Themen wie Kontext, Parallelität und Fokus im agentischen Arbeiten so stark ins Gewicht fallen. Spannend wird es dort, wo das Konzept weitergedacht wird: Mit Semantic Contracts beschreibt Ralf eine Möglichkeit, auch Begriffe für Wissen außerhalb der Trainingsdaten definitorisch zu fixieren. Daraus entsteht ein Workflow, der von Requirements über Spezifikation und Architekturdokumentation bis in eine Dark-Factory-artige Umsetzung reicht. Zum Schluss öffnet die Episode noch den Raum für eine größere Frage: Wie kann Vertrauen in KI-Systeme wachsen, wenn klassische Reviews nicht mehr skalieren und stattdessen Fehlerkorrektur, Compilerläufe und andere Rückkopplungen den verlässlicheren Kanal bilden sollen? -- Links und Referenzen arc42: https://arc42.org/ Semantic Anchors: https://rdmueller.github.io/pages/blog/semantic-anchors.html Semantic Contracts: https://llm-coding.github.io/Semantic-Anchors/contracts Eichhorst Prinzip: https://rdmueller.github.io/pages/blog/eichhorsts-principle.html Ralf D. Müller: https://rdmueller.github.io/index.html Der HMZE Podcast ist ein Projekt von Sebastian Heide-Meyer zu Erpen und André Neubauer. Die Inhalte stammen von uns und unseren Gästen. Weiterdiskutieren kannst du bei LinkedIn oder besuche uns auf unserer HMZE Seite, wo wir alle Folgen veröffentlichen. Fragen und Anfragen kannst du uns gerne per Mail an ⁠⁠ schicken. Vielen Dank für deine Zeit und bis zur nächsten Folge.

    • Transcript
Showing 1–20 of 22 episodes