Show Notes
How do you prove AI is shipping more features? Amos Haviv leads the Developer Workflow teams at Booking.com, supporting 4000 engineers operating 8000 repos. Everybody is burning through their AI budget right now and almost nobody can answer what it bought them. Amos can, because his team spent four years building an event system to debug their own SDLC before AI upped the urgency. In this video, we cover: Why verification is the bottleneck right now, and where it moves next Building an event store that separates KTLO from real feature delivery Why static dashboards create the metric they measure, and the cobra story behind it Agent cost, model routing, and why Booking ignores token maxing entirely Running a developer survey with a 92% response rate across 3k+ engineers Who should own skills and MCPs: a central platform team or the domain experts? For platform engineers, engineering leaders, and anyone being asked to prove ROI on AI tooling this quarter.
Timestamps: 00:00:00 - Everyone is burning through their budget 00:00:32 - Verification Is the Bottleneck Every Team Hit 00:03:35 - 4,000 Engineers and 8,000 Repos at Booking.com 00:06:48 - Why Copying Google and OpenAI Will Break You 00:09:21 - Verification Is a Stack of Agents, Not One Review 00:13:27 - Cost Is Becoming a Bottleneck of Its Own 00:17:14 - Was the Internet a Bubble? What That Teaches Us 00:25:32 - What Working With the Frontier Labs Looks Like 00:28:26 - Debugging the SDLC With Four Years of Event Data 00:30:24 - Do Engineers Using AI Actually Ship More Features? 00:37:13 - Where to Start If You Measure Nothing Today 00:45:01 - The Cobra Effect: When a Metric Becomes a Target 00:52:23 - Everyone Is a Builder Now, and Everything Needs Support 01:01:21 - Is AI Turning Every Engineer Into a Manager? 01:03:46 - The Developer Survey With a 92% Response Rate 01:10:09 - Who Owns Skills, MCPs, and the Enterprise Harness 01:17:46 - Great Developer Experience Is High Velocity
Mentioned in the episode: High Output Management by Andy Grove The Sovereign Individual (1997) The story of General Magic
Views expressed are Amos's own and do not represent Booking.com. #AI #SoftwareEngineering #DeveloperExperience
More from Beyond Coding
How do senior engineers use coding agents on prod infrastructure when “make no mistakes” is just a polite request? Kerim Satirli (HashiCorp, an IBM company) explains why prompt rules don’t hold in production, and what he runs instead. In this video, we cover: How Infrastructure as Code (Terraform, Contentful, Datadog, Okta, Cloudflare providers) evolved from open-source PRs to automated agent harnesses Why AI skills files and “make no mistakes” prompts are only requests, and what replaces them in production pipelines How AI-generated pull requests are overwhelming open-source maintainers (and why projects like cURL are pushing back) Personalized “software factories”, spec-driven development (GitHub Spec Kit), and cutting AI documentation noise with BLUF (Bottom Line Up Front) What open-source Rust clones of Adobe mean for the future of SaaS For infrastructure engineers, DevOps practitioners, and software architects who want to move past one-shot prompting and run coding agents on production infrastructure without losing control. Timestamps: 00:00:00 - Why “Make No Mistakes” Is a Polite Request 00:00:38 - How Open Source Terraform Changed Provisioning 00:01:59 - Why API-First Services Matter Beyond AWS and Azure 00:04:19 - How AI Agents Changed Terraform Provider Updates 00:07:03 - Why “Make No Mistakes” Won’t Stop Your Agent 00:10:31 - Bringing Determinism Back to Agentic Workflows 00:13:07 - How AI Pull Requests Are Overwhelming Open Source 00:15:41 - Why Every Engineer Is Building a Software Factory 00:18:52 - Why AI-Generated Documentation Slows Down Humans 00:23:12 - Using Agents to Filter Noise and Summarize Specs 00:27:02 - Escape Hatches for Automated Terraform Applies 00:29:03 - Where the Joy in Engineering Lives Now 00:31:24 - Spec-Driven Development and the Return of Architecture 00:34:51 - Open-Source Rust Clones, Reverse Engineering, and SaaS 00:38:19 - How to Build Software Products People Actually Want
Lead Architect Carlos Kelkboom shares his rules for coding agents and how great engineers architect software with AI. One of the great ones is: he throws away generated code the moment he doesn't understand it. He breaks down why coding agents fail at iterative changes, why his main agent never touches files, and how he scaffolds projects for one-shot execution. In this video, we cover: Why AI coding agents struggle with iterative changes (SlopCodeBench) and excel at one-shot execution How to manually scaffold code, types, and folder boundaries before letting sub-agents run Designing a distributed Markdown and indexed-search graph harness for progressive context disclosure Using Git worktrees to run parallel coding agents and merge the best implementations Why dynamic memory and the heap need to disappear in favor of deterministic programming languages How AI bubbles and solo coding impact junior engineers and team collaboration This episode is for software engineers, tech leads, and system architects who want to move beyond "vibe coding" and build maintainable, deterministic production software with AI agents. Connect with Carlos: https://www.linkedin.com/in/carlos-k-4016b8225
What happens when coding agents see your entire system: code, services, infrastructure and live traffic, all connected? Most agents are blind to everything outside the repo. Adyen Staff Engineer Mat Jones explains what changed once theirs weren't, and why he calls it somewhat scary. In this video, we cover: Running five or six coding agents in parallel with git worktrees and Atrium, an internal tool that lets agents coordinate and share memory Architect: nine connected graphs in Neo4j covering services, code, DNS, data stores and infrastructure Ingesting 7 million trace spans a second with Kafka and Apache Flink, and what collapsed on the first try Sentinel: anti-pattern monitoring, with agents that raise the fix as a merge request MCP vs APIs vs CLI tools, and how to deal with tool bloat This episode is for software engineers and platform teams who want coding agents to reason about real systems, not just the repo in front of them. You'll leave with a concrete place to start and an honest view of the risks.
HumanLayer 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 trust Token maxing, dark factories and the "Instagram phase" of AI engineering Why 2-4x productivity beats chasing 100x Pull requests, design docs and how engineers build taste now Dex's agentic engineering workflow: one-shots, overnight janitor bots and research-design-plan For engineers and engineering leaders adopting AI coding agents who want to know what's real and what works in teams at scale. Timestamps: 00:00:00 - The Man Who Coined the Term: Context Engineering 00:00:54 - SlopCodeBench: Proof AI Makes Codebases Worse Over Time 00:05:49 - Background Agents and the Enterprise Friction Problem 00:08:13 - Why You'll Never One-Shot a Big Feature 00:11:18 - Scaling Agentic Engineering to 1,000 Engineers 00:16:19 - Dark Factories, Fake Hype and Token Maxing 00:19:55 - Why 2-4x Faster Beats Chasing 100x 00:21:31 - Should Teams Stop Doing Pull Requests? 00:24:20 - Can Deterministic Tools Replace Code Review? 00:27:37 - How to Build Engineering Taste When AI Writes Code 00:33:05 - Dex's Agentic Workflow: One-Shots and Janitor Bots 00:35:18 - Research, Design, Plan: Shipping Big Features With Agents 00:38:35 - Why Overnight Agents and Heavy Parallelism Backfire 00:40:10 - Subscription Maxing vs Paying Per Token 00:42:40 - Parallelize Like Poker, Not a Slot Machine Dex Horthy: https://x.com/dexhorthy Dex's talks and deep dives: https://aithatworks.bold.video #AgenticEngineering #AICoding #SoftwareEngineering
Amp ships to production around 50 times a day with AI agents, but the change that makes their team ship faster isn't AI at all. Dario Hamidi, AI engineer at Amp, makes the case that pull requests are holding your team back, and explains what replaces them. In this video, we cover: Why Amp works without pull requests or feature branches, and how to try it with your own team Agent permissions, approval fatigue, and building environments where you can trust agents Orbs: remote agent execution, parallelism, and multiplayer sessions Short-lived credentials and workload identity federation for cloud agents Triaging 60+ bug reports a day with agents without frying your brain For engineers and engineering leads who want to ship faster with AI agents and are ready to question the workflows they've always used. Timestamps: 00:00:00 - Is It Time to Kill the Pull Request? 00:00:45 - What Happens If Every PR Gets Auto-Closed Tomorrow 00:05:04 - Pull Requests Were Built for Strangers, Not Teammates 00:07:20 - Does Dropping PRs Hurt Quality at Scale? 00:08:53 - Agent Permissions: Why No Default Makes Everyone Happy 00:12:42 - Don't Juggle Knives: Building Environments You Can Trust 00:14:56 - The Hidden Hard Parts of Cloud Agents 00:16:39 - What's an Orb? Remote Agent Execution Explained 00:19:10 - How to Give Agents Credentials Without Leaking Secrets 00:22:51 - Why Orbs Live Forever (and Multiplayer Orbs) 00:25:49 - The State Management Problem Nobody Talks About 00:29:22 - When Will Everyone Have Remote Dev Environments? 00:32:23 - Orbs That Spawn Orbs: Orchestrating Agents at Scale 00:34:42 - How Dario Triages 60 Bug Reports a Day With Agents 00:39:23 - Avoiding Burnout When Agents Remove the Ceiling #AIAgents #SoftwareEngineering #PullRequests