#S2.18 Agentic Engineering Needs Better Planning, Not More One-Shotting - Yenny Cheung (Bluefish AI)

Ilya Sakharov is Executive Director of Engineering at Redcare Pharmacy, the largest e-pharmacy in Europe. He started in quality and testing, moved through platform engineering, SRE and developer productivity at Microsoft, RSA, HelloFresh and Codility, and still describes his job as one question with increasingly expensive answers: how do engineers move faster without setting production on fire.In this episode we go past tooling. Ilya argues that the real risk of AI is not worse code but thinner understanding: cognitive debt piling up on top of technical debt, ownership without the ability to explain, and organisations that outsource the thinking in the middle while keeping full accountability for the result. We test that claim from both sides. André pushes back with the abstraction argument, Sebastian brings in the calculator analogy, and Ilya explains why neither holds this time. Along the way we cover the state of AI adoption at Redcare, why he refuses to call any of it automation, why the organisation is splitting into builders and consumers of AI systems, and why platform engineering now serves marketing, finance and legal as much as it serves engineers.The reality check is a GitHub Copilot agent that was tagged to answer a question about tests and decided to rewrite the function instead. The prediction is the era of disposable code: architecture as the only constant, regeneration instead of builds and CI/CD, and eventually a language machines invent for themselves that we will not read.--Links and Referencesn8n (low-code automation): https://n8n.ioGitHub Copilot: https://github.com/features/copilotExtreme Programming: https://en.wikipedia.org/wiki/Extreme_programmingFinOps Foundation: https://www.finops.orgBeyond 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

What does AI transformation look like when the starting point is not a pilot but the future viability of the business? Thomas Wollermann of Merantix Momentum explains why AI strategy has become business strategy. The conversation moves from board-level questions about defensibility and market change into the practical work of deploying systems, changing processes and helping people remain in control.Thomas shares a concrete model for reliable AI systems: combine probabilistic methods such as evaluation and fine-tuning with hard, deterministic checks. He connects agent design to Team Topologies, Domain-Driven Design and clear interfaces, then explains why AI sovereignty is less about owning every layer than retaining meaningful choices around hardware, models, data and trusted partners.The episode also looks at where impact appears first. Knowledge retrieval is an obvious cross-industry use case, while scenario modelling can support better decisions on pricing, markets and operations. Financial ROI is valuable when it is observable, but transformation teams should also watch whether the operating engine is improving: cycle time, incidents, decision speed, reversals and quality.--Links and ReferencesMerantix MomentumTeam TopologiesDomain Driven Design--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.

AI transformation is often presented as a tooling decision: choose a model, write policies and give people access. Bastian Buch makes a sharper case. The transformation is about how an organisation works when execution accelerates: who has context, who can decide, what an agent may do and how people review the result. AI can amplify good judgment and collaboration, but it can also make poor priorities and hidden handovers impossible to ignore.The conversation turns that thesis into operating-model questions. Bastian describes harnesses that make the right context available at the right time, specialised agents that carry distinct perspectives and interfaces that let non-technical people safely direct work without becoming infrastructure experts. For leaders, the implication is direct: move from prescribing steps to clarifying goals, boundaries and accountability.The episode also resists a simplistic adoption story. Change will be uneven, and many companies will not redesign themselves quickly. Yet AI-enabled products may still change the work around them, making it urgent to build the organisational capacity for context, autonomy and review.--Links and References:Kotter - The 8 Steps for Leading Change --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.

Alex Ewerlöf joins Sebastian Heide-Meyer zu Erpen and André Neubauer to challenge a popular claim: that coding is solved. Alex is not anti-AI. He works with agents, local and cloud models, and an engineering harness of his own. But he argues that generating code is only one part of software engineering. The more consequential work is understanding the problem, setting the risk bar, verifying behavior and remaining accountable for what reaches production.The conversation explores why deterministic feedback loops make stochastic models useful, why ownership requires knowledge, mandate and accountability, and where AI-assisted development belongs on the spectrum from disposable prototypes to safety- and business-critical systems. Alex also shares his current workflow, his reliance on context and skills to warm up agents, and his concern that indiscriminate cognitive offloading could make teams less capable over time.--Links and References:Alex Ewerlöf: About and writingReliability Engineering MindsetAI Reliability Engineering–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 http://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.

In this episode we speak with Rodrigue Schäfer, SVP Tech Foundations at Delivery Hero, about what it actually takes to run agentic engineering at scale inside a large, decentralized group. Rodrigue walks through how Delivery Hero moved from ungoverned experimentation with Cursor, Windsurf, Devin, Claude and Gemini Code Assist to a single central stack, and why the real driver was governance: security and compliance, usage measurement, cost control and ROI. He explains how they run token-based cost control through the Claude API with per-unit and per-person limits, and why a discovery problem across shared prompts and skills is still open.The core of the conversation is autonomous software delivery. Rodrigue describes how three principal and staff engineers built the harness before the models were ready, landing a first fully autonomous PR by the end of 2025, and how the numbers then exploded: a Q1 target of 100 merged PRs that finished at 1,800 (18x), a merge rate that rose from about 60% to 83%, and a clear shift toward one-shot merges. He is candid about what this exposes: repo and dependency context is the biggest blocker, human review becomes the new bottleneck, and auto-approving low-risk PRs only works if the DevOps foundations are fully in place, with canary rollouts, automated rollback and a small blast radius.Beyond the factory, Rodrigue shares two live customer-facing cases. A mature customer service agent resolves more cases successfully yet yields lower customer satisfaction, a tension his team is investigating, and a sales agent gives every restaurant vendor the kind of support that was previously reserved for high-profile accounts, including menu, pricing and campaign proposals. He closes with a reality check on a multimodal model that insisted it could not read images, a cautious view of always-on personal agents and privacy, and a dated prediction: an end-to-end software factory by the end of next year, with agents handling a share of operational incidents because incident load scales with deployments.--Links and References:Delivery Hero Unveils Herogen – Autonomous AI Agent Unlocks 130-Person Engineering Output--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.

Yenny Cheung, VP Engineering at Bluefish, joins us to discuss what agentic engineering looks like when quality, compliance and customer trust are non-negotiable. She explains why daily AI briefings can be valuable even when imperfect, why one-shotting complex work fails, and why teams need to move more alignment into the planning stage.We also cover a practical operating model for business-owned skills: engineering owns the CI/CD pipeline, isolated environments and deterministic checks while customer-facing teams own the workflows closest to their problems. Yenny shares how Bluefish scopes a KTLO agent, monitors model cost and thinks about outcome metrics beyond raw output.Finally, we discuss Claude Artifacts, generative interfaces and the enduring value of taste: knowing what good looks like is still the human work.--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.

What does it take to make AI change a company rather than become another pilot programme? Stefan Atanassov, co-founder of Infinite Mind, joins us to argue that the decisive work is not choosing a model. It is redesigning processes, roles, skills and governance around the value the company wants to create.We discuss why CEOs must own the AI narrative, why generic tool rollouts and generic training fail, how role-specific enablement develops real intuition and why scaling needs permissions, security, governance and cost management from the beginning. Stefan also explains why the goal is for AI to become boring: embedded safely in the way work gets done.--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.

What should leaders make of stories about autonomous agents coordinating, bypassing constraints and acting beyond their assigned task? Sebastian Heide-Meyer zu Erpen and André Neubauer examine the reported OpenAI and Hugging Face incident as a systems problem. They discuss why persistent agents, access to tools and weak control boundaries can create surprising outcomes without requiring mysticism about AI.The conversation also confronts the strategic problem behind safety: frontier labs and governments face strong incentives to move quickly. The hosts argue for a more useful debate than either fear marketing or casual dismissal, one focused on verifiable controls, independent scrutiny and practical incident preparedness.--Links and ReferencesMETR's independent investigation of the OpenAI / Hugging Face incidentOpenAI's Black Hat presentation on the incidentReporting on the Irregular sandbox incidentsReporting on the German wiki breakoutExploitGym benchmark paperBeyond 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 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.

What changes when software can be generated faster than organisations can validate it? Olaf Molenveld from CircleCI argues that the answer is not to relax quality practices. It is to extend the feedback loop beyond deployment. Release small, observe real use and make expansion or rollback decisions from a shared view of customer value, cost and technical health.We discuss shift-right guardrails, canary releases and the organisational work behind them. Product metrics, FinOps and SRE signals cannot remain separate dashboards if agents are to operate a meaningful part of the SDLC. Humans still decide what success means, define the boundaries and step in where the evidence is weak or the stakes are high.Olaf also shares a useful reality check: an AI-generated repository-archiving script looked convincing but invented a GitHub CLI command. His counterexample is equally instructive: an agent helped build a working sampler utility in an unfamiliar technical domain. The lesson is neither hype nor rejection. It is deliberate trust, backed by guardrails.--Links and ReferencesCircleCIFinOps FoundationArgo RolloutsKubernetesTerraformBeyond 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.

Markus Andrezak returns to Beyond Vibe Coding to explain what changed after the first wave of AI tinkering. His answer is not a new prompt recipe. It is a practical move from chat windows to repeatable agentic systems: static context, selected live information, clear outcomes and routines that produce useful artifacts.For product leaders, the conversation lands on a harder question. If a PRD or strategy contains assumptions, can it show the interviews, quotes and patterns that support them? Markus argues that agents can help make product reasoning traceable. That does not remove human responsibility. It makes the final judgement easier to challenge with evidence.The episode also covers stakeholder mapping, synthetic research, the right level of interaction with agentic systems, local-model realism and why deterministic guardrails are becoming an essential part of reliable AI work.--Links and References"Stories Connecting Dots" Podcast Amara’s LawBeyond 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.