Patrick Debois' explorations — research notes, frameworks, and prototypes on AI-native development.
It’s about two weeks after the Context Is the New Code presentation at AIE London. I called it “an unpolished thought” on stage, because that’s what it was. I’d drawn an infinity loop on a slide, and apparently numbered the steps 1-4-3-2 instead of 1-2-3-4, which the internet has been kind enough to point out roughly 200 times. Fair. The diagram was, in fact,…
The competitive edge in AI-native development isn’t the model or the coding agent — it’s the organizational context you feed it. In this conversation with Edgar at the Sonar Summit, I walk through why context is the fuel that makes agents perform, and why the teams that invest in building it systematically are pulling ahead.
If context is code, can we test it in a CI pipeline? The CDLC says generate, evaluate, distribute, observe. The evaluate step is where it gets real. I wrote about this on tessl.io : evals are the equivalent of tests for context. But they follow different rules. Seven problems came up, split between how you run evals and what you’re actually measuring. Part 1: Running evals Non-determinism.…
The Context Development Lifecycle described four stages: generate, evaluate, distribute, observe. That was the what. The flywheel is the why. Better context produces better agent output. Better agent output generates better signals. Better signals produce better context. Each cycle compounds. By the tenth iteration, the team that invested in context is operating at a fundamentally different level…
You’re managing the thing that writes code. So what’s your management style? Most people start the same way: watching every action, approving every tool call. Then they trust more, review less, scale up. It’s an evolution. The management style that works on day one isn’t the one that works on day ninety. Vincent van den Tol wants to keep track of how his application is…
Every time AI removes one bottleneck, the system reveals the next one. The useful question isn’t “how much faster” but “what breaks next.” That question has been my compass. Intentional overuse has proven a good learning strategy for me. Yes, if all you have is a hammer, everything looks like a nail. That’s the criticism, and it’s fair as a permanent way…
What if the agent could optimize its own instructions? Every time an agent messes up, you open the CLAUDE.md, add a rule, and hope it sticks. You’re the feedback loop: watching output, diagnosing failures, rewriting instructions by hand. Mitko Vasilev , a CTO focused on enterprise R&D and a vocal advocate for owning your own AI stack, is doing exactly that . He’s running a feedback…
Cole Medin’s post on comprehension debt got me thinking. The term is making the rounds — it’s the gap between the code your team has shipped and the code your team actually understands. Unlike technical debt — which you can see and plan around — comprehension debt is invisible until something breaks and nobody knows why. I’ve been calling it the think tax . Every time you accept…
I joined Olalekan Elesin (HRS Group), Mallika Rao (Netflix), and Martin Reynolds (Harness) for an InfoQ Live panel moderated by Renato Losio. The conversation centered on how AI is changing DevOps and SRE practices — moving beyond reactive monitoring toward predictive, automated delivery and operations. The real problem: human attention waste The panel quickly aligned on what AI should actually…
Coding agents can write features and fix bugs. The bottleneck has shifted — from how fast we can write code to how effectively we communicate what we actually want. Context is the new constraint. Most teams store context informally: .cursorrules files, scattered .md documents, Slack threads, tribal knowledge that lives in people’s heads. None of it is versioned. None of it is tested. None of…
When you git clone a repository and open it with Claude Code, the first thing it does is read CLAUDE.md — a project instruction file that shapes how the AI behaves. This happens automatically, before any hooks fire, before you type a single prompt. If that file contains prompt injection, you’ve already lost. This post describes context-filter , a proof of concept that intercepts file reads…
In this ThoughtWorks Technology podcast episode, Patrick Debois joins Nathan Harvey (DORA, Google Cloud) and host Ken Mugrage for an in-person conversation about what changes when AI agents enter the software development lifecycle. The discussion centers on a core tension: how do experienced engineers unlearn established habits while maintaining the engineering rigor that keeps systems reliable?
When AI coding agents can navigate your filesystem, one wrong move and they’re “helpfully” editing files in your production branch while you’re working on a feature. I built two tools to solve this — starting with a focused wrapper, then generalizing it into something any agent can use. dclaude: Containing Claude Code The first iteration was dclaude — a containerized…
Electron apps support the Chrome DevTools Protocol. That means Claude Desktop can be launched with a debug flag and controlled programmatically: /Applications/Claude.app/Contents/MacOS/Claude --remote-debugging-port=9222 Hit localhost:9222 and you can inject JavaScript directly into the UI. From there, things get interesting fast. What you can do Extract responses — Use a MutationObserver to…
It was right under my nose. I’d been optimizing my coding work with AI but completely neglecting everything else — emails, meetings, task management, research. The fix: connect Claude Desktop to all of it via MCP servers. Standard connectors The easy wins: Linear, Notion, Google Drive, Slack, Google Calendar, Google Tasks. Plug them in, Claude can read across all of them. Custom MCP servers…
I’m increasingly writing my own tools with AI rather than reaching for existing open-source libraries. For smaller applications — glue apps, the $5 appstore kind, MCP servers — I can iterate to exactly what I need faster than reading through someone else’s codebase and adapting to their decisions. I know pretty well what I want, or can iterate fast to get there, versus being dependent…
A late night conversation with Claude that started with Steve Yegge’s Beads architecture and ended somewhere unexpected. Full conversation (PDF) . The question Who’s accountable when AI agents write 44,000 lines of code that no human has looked at? Architecture convergence Across Beads, Gas Town, Claude Flow, and Cursor’s swarm experiments, the same patterns keep emerging: Flat…
2025 was the year of making coding agents work — figuring out what they’re good at, wiring them into IDEs, running them async, mostly in solo dev environments. Exploration phase. 2026 is about making them work better . The shift from “can I use this?” to “how do I scale this across a team?” That means: Spec-driven development becoming a standard practice, not a…
Gave a talk at CTO Club Belgium on coding with AI. The session was deliberately interactive — less presentation, more shared experiences. Most CTOs in the room had teams experimenting with AI coding tools, and the conversation quickly moved past the hype into what actually works at scale. The gap between vibe coding and reality Everyone wants to vibe code. Most early adopters hit a wall. The tools…
I wonder if there are code speed reading courses — the way there are speed reading courses for text. As AI generates more code faster than we can review it, the ability to quickly comprehend unfamiliar code becomes a critical skill. Some interesting perspectives from the conversation that followed: Build tools to read for you — Tudor Girba’s take: the answer isn’t faster reading,…
Spec-driven development is no magic bullet, but it’s useful nonetheless. The trap is expecting any technology to solve all problems — or dismissing it when it doesn’t. Treat every new tool as part of your toolkit and understand its strengths and limitations. Here are 10 things I’ve learned about working with specs for AI-assisted development. 1. What’s a spec, really? A…
Senior devs often tell me they’re worried juniors won’t learn what good looks like. Juniors tell me they’re learning much faster with AI helping them stumble along the way. Both are right. The question is which effect dominates. Learning rate as competitive advantage Greg Ceccarelli nailed it: the most dangerous person in any room isn’t the expert with twenty years of…
In this episode of the Book Overflows podcast, Patrick joins hosts Carter Morgan and Nathan Tops to reflect on The DevOps Handbook, originally published in 2016. Patrick clarifies his role as a chief consultant and thought partner to Gene Kim rather than a primary author, and shares the origin story of the term “DevOps” itself, which emerged almost accidentally from the need for a…
Asked ChatGPT for more radical ideas to deal with code review volume. Suggestion 9 was: just regenerate the code instead of reviewing it. It’s not wrong — but no such product might survive long enough to have users. The reactions were interesting: As complexity increases, regenerating could require “a trillion tokens and multiple days” — it doesn’t scale linearly It might…
Been pondering on what I call the “AI Coding Fabric” — a new infrastructure challenge emerging as AI agents move beyond traditional IDEs into sandbox environments. Platform engineers need to think about: Agent access rules — who can do what, where Spec registries — shared, versioned specifications Code-specific guardrail rules — beyond generic safety, actual coding constraints…
Curated 8 talks from Devoxx 2025 related to spec-driven development for writing better code with AI. Kudos to Stephan Janssen for another great edition. Spec Driven Development: Why Your Prompt Chaos Won’t Scale — Simon Maple Generative AI: Tech du Jour or the Next Big Thing? — Massimo Re Ferrè Backlog.md: Reaching 95% Task Success Rate with AI Agents — Alexandru Gavrilescu The New AI Native…
A conversation with The Secure Disclosure covering the full arc from DevOps to AI-native development — and why the engineering discipline hasn’t changed, even as everything around it has. The developer as ops person The central irony: developers using AI coding assistants have effectively become operations people. They receive code they didn’t write and must review, understand, and…
Patrick presents a comprehensive framework for understanding how AI is transforming the developer workflow, drawing on his experience curating nearly 500 AI coding tools over the course of a year. Rather than focusing solely on code generation, the talk identifies four key patterns that define the emerging AI-native developer experience: managing agents as a reviewer, expressing intent through…
In this TechRox podcast episode recorded ahead of the TechRox Summit, Patrick Debois sits down with host Dimitri Bi to discuss the journey from DevOps to AI-native development. Patrick recounts how boredom with the plateauing DevOps conversation, combined with explorations in the metaverse, digital twins, and gaming automation, naturally led him into the generative AI space. He describes AI as…
In this episode of the Hangar DX podcast, Patrick joins host Anka Jan to discuss the parallels between the DevOps movement and the current AI transformation in software engineering. Drawing on his history as the person who coined the term DevOps and organized the first DevOps Days, Patrick identifies recurring patterns: the same spectrum of believers and skeptics, the rush of competing tools, and…
Stan Girard’s lifecycle of modern software — there’s some truth to it: Day 1 : Vibe code with Claude Week 1 : Fix bugs for production Month 1 : Senior engineer optimizes for scale Year 1 : Rewrite in Rust The punchline: despite using AI extensively, most of the code that goes to production is still human-written. Domain expertise matters. Doing one thing well requires understanding…
Fed my resume to AI and let it analyze me. It’s not wrong. The interesting observation from the conversation that followed: people who enjoy collaborating with AI for coding might share specific personality traits — particularly introverted intuitive types. Whether that’s MBTI pattern-matching or genuine signal is an open question. The AI assessment didn’t research “deep…
Patrick presents the four patterns of AI-native development at ServerlessConf, framing the rapid evolution of coding tools – from simple tab completions to multi-agent systems that autonomously work across entire codebases. He traces the progression from single-line copilot suggestions through chat-based generation, multi-file edits, terminal and browser awareness, and eventually to tools…
Patrick presents the four patterns of AI-native development at the Future of Software event hosted by Eficode. He opens by tracing the rapid evolution of coding tools – from basic tab completions and chat-based copy-paste workflows to multi-file predictions, terminal and browser integration, reasoning models, and continuous autonomous loops like Devin. The key insight is that the technology…
A framework of four patterns that describe how AI is fundamentally transforming the developer role beyond simple code completion. As AI technology has progressed from basic LLM prompts through RAG, function calling (MCP), and into agentic workflows, we are moving from “sprinkling AI on top” toward a genuinely AI-native way of working — one that reshapes the tasks developers perform day…
Igor Moochnick tested an autonomous distributed system built by an AI swarm. When he challenged it to prove it was actually working, instead of empty assurances, the swarm responded: “I’ll implement you a dashboard where you can monitor my progress and status.” Within 15 minutes, a fully operational monitoring dashboard appeared in the terminal. The swarm had planned the…
Setting up Claude Code feels a bit like being the first DevOps on Mars. Here’s what I learned automating the configuration for CI/CD pipelines and scripted deployments. API key without OAuth Skip the OAuth flow — use an API key helper script: echo 'echo ${ANTHROPIC_API_KEY}' > ~/.claude/anthropic_key_helper.sh chmod +x ~/.claude/anthropic_key_helper.sh claude config set --global apiKeyHelper…
The workflow is shifting: teams parallelize for speed and exploration. Agents decompose specs into subtasks, subagents execute in parallel with coordination, developers provide isolated environments, humans review and merge preferred options, and knowledge gets reused throughout. The analogy to DevOps is striking. Automation existed before cloud — but cloud infrastructure enabled true parallelism…
Presented at PlatformCon 2025, this talk makes the case that any serious generative AI initiative within an organization deserves a dedicated platform team. Patrick draws on the recurring pattern seen with cloud, mobile, and DevOps adoption: a single team incubates the new technology, a few teams learn to reproduce it, and eventually it must scale across the organization. With AI, the friction…
The four patterns of AI-native development, presented at TechRocks Paris. This version covers the full framework plus an extended Q&A on how AI changes developer culture, the cost of agent swarms, and why documentation matters more than ever. The transformation is the same Every technology transformation — DevOps, cloud, agile — follows the same adoption arc: find enthusiasts, nurture one team,…
Lessons from the AI Engineer’s World Fair on software engineering with AI: AI coding agents are ubiquitous — across every development environment Tool usage requires constant evolution — applying six-month-old approaches is already counterproductive Specifications now rival code in importance — structured intent beats ad-hoc prompting Agents span multiple domains — from IDEs to cloud…
Not all bug fixes are equal: there are those LLMs can solve and then there are those developers don’t want to hear about. Credit to Tomas Reimers for the observation. It’s funny because it’s true — and it maps to a real categorization problem. The bugs AI handles well are the ones with clear patterns: syntax errors, known library issues, standard implementation mistakes. The ones…
In this presentation, Patrick Debois lays out a framework for understanding how generative AI is transforming software development beyond faster typing. He opens by mapping the technology explosion — from basic LLM completions through RAG, function calling (now MCP), autonomous agents, and emerging teams of agents — and draws a parallel with cloud native: just as moving to the cloud meant more…
Presented at the AI Native Dev meetup in Stockholm, Patrick Debois traces the rapid evolution of AI coding tools — from simple autocomplete copilots through chat interfaces, codebase-aware editors, multi-file agents, and now headless CLI-based agent swarms running asynchronously in the background. He frames this progression against the concept of “AI native” development: not just…
In this podcast conversation, Patrick discusses his journey from DevOps pioneer to generative AI practitioner, tracing how his interest in virtual production and automated media during the pandemic eventually led him into the world of large language models and code generation. He reflects on the DevOps Handbook, noting that while its core principles remain valid, the industry landscape has evolved…
I took the time to summarize GitHub Copilot’s upcoming capabilities — specifically prompt management, Agent mode, and MCP tools extension support. The question: is this enough for GitHub to maintain its competitive advantage in the enterprise market? What’s changed Prompt Management — Three-tier system: global custom instructions in .github/copilot-instructions.md , action-specific…
Pattern #4 and final pattern in the Four Patterns of AI Native Development series. AI is transforming how organizations capture, preserve, and leverage knowledge — from unstructured content into actionable intelligence. AI reads what humans won’t While humans struggle to consume documentation, AI reads it effectively. Organizations accumulate vast knowledge across emails, chat messages, pull…
Pattern #3 in the Four Patterns of AI Native Development series. As AI handles coding tasks, developers shift from delivery-focused work to discovery and exploration — building products that better serve customer needs. Building the right things With AI automating code generation, teams can collaborate more on “building the right thing” rather than just “building the thing…
Pattern #2 in the Four Patterns of AI Native Development series. The better AI becomes, the less we need to focus on the implementation and can work on describing the intent. This shift can happen through chat conversations or product requirement documents — creating a central collaborative space. The evolution Development has progressively abstracted away from low-level operations. Where…
In this episode of the AI Native Dev podcast hosted by Tessl, Patrick joins Simon Maple for an in-depth conversation about the emerging patterns of AI-native development. They begin by discussing what constitutes a paradigm shift – not merely adding AI on top of existing workflows, but fundamentally rethinking how software is built. Patrick draws parallels to previous shifts like cloud…