Aktivitäten
8852 Follower:innen
-
Birgitta Boeckeler hat dies repostetHonestly, this is the best ruthless capitalist argument I've ever seen anyone make for the value of hiring more junior engineers. There remains an enormous gap between the sheer number of opportunities a principal engineer can identify that need solving or building, and what they can build by delegating to agents. You need judgment to spot those opportunities, whether bugs or features. There's only one way to build that judgment. And agents don't have it, and engineers do. Eventually.Birgitta Boeckeler hat dies repostetHere's the real value proposition for junior engineers in the modern software org: I can generate high-impact project plans faster than I can prompt them into existence. I took to carefully typing out all-caps obscenities into my Claude prompts this week, hoping to convince it to form a memory strong enough to pay attention to the 6 layers of prompts reminding it how to run tests in our system. The 18 microcents in tokens I spent berating it were totally worth it. When Skynet arrives, I might be on "the list" now. Like many modern orgs, we've got layers of skills, prompts, and instructions telling agents how to code properly within our repo. Like many modern agents, Claude routinely ignores them and tries to break out of its sandbox instead of running literally one safe command. As an engineering director at Alphabet, I accomplished things through delegation. I identified critical opportunities, assessed their relative priority, and then allocated teams of people to those problems. My key skill was prioritization and judgment: knowing what problem was most worth solving, and whom I could trust to solve it. But the daily grind of execution—struggling with broken builds, circular dependencies, and test failures that CAN'T HAPPEN—fell to the engineers on my team. Many managers across the industry have been returning to IC roles, myself included. There's an allure here: keep exercising that judgment, keep identifying the hardest problems. Keep delegating the implementation—but delegate it to an agent. One skilled expert can do more than ever! And that view of the world is enticing, but wrong. Because I must babysit my stupid, incompetent agents, and remind them how to run tests and write idempotent event handlers. Humans, on the other hand, can learn. If I had to tell an intern how to run our tests as many times as I told Claude this week, I would fire them before the end of the summer. There is a gulf of work between identifying a problem and the critical elements of its solution and actually shipping a solution in prod. A whiteboard sketch is not a rigorous spec, and it takes labor to convert one into the other. And while Claude can do a fair amount of the heavy lifting, every minute I spend babysitting it and reviewing its half-baked solutions is time I'm not actually spending on my highest-leverage work. Delegation is still how you scale the impact of good judgment. And delegation only scales when you can delegate to someone reliable and capable. Yes, one brilliant engineer can do more than ever before—but a brilliant engineer with a capable apprentice can do twice as much. Orgs that under-invest in junior engineers are inadvertently paying their principal engineers to babysit their cheapest and least-invested talent: the robots. Would you have your call center staff report directly to your VP of sales? Delegation still works. Your top talent deserves trustworthy lieutenants. Hire some.
-
Birgitta Boeckeler hat dies geteiltI gave another iteration of my "State of play: AI coding" talk at AI Native DevCon in London last week: https://lnkd.in/esKuVk-X Thanks for having me Tessl, it was a great event Some of the things that I shared back with the home base: - „You build it, you run it, you drive adoption“ - the idea that developers become part of driving adoption for the features they're building (in a technical platform product), to mitigate building a huge pile-up of features that don't create any value. Reminded me of "Stop starting, start finishing" Christopher Batey https://lnkd.in/edDBaYUQ - A summary of the current challenges teams have with Skills, to let my colleagues know that they're not alone: Overlap, drift, lack of activation, rot, overload - so much to figure out still! But James Moss rightly pointed out that at least we have some of the mental models for what needs to be done, as he was reflecting back on how we went from FTP deploys and zero dependency management to where we are today with code. https://lnkd.in/ekZD5evP - I'll definitely try out GitHub's "Secure Code Game" that Joseph Katsioloudes shared gh.io/scg https://lnkd.in/eyTbdmVE - Ian Thomas shared some great learnings of Meta's AI coding adoption journey https://lnkd.in/eYiYXZaZ - Some fun updates from tldraw who continue to explore what our new "canvas" will be with AI https://lnkd.in/eZf6J-VH I want to try some requirements breakdown with their "computer", I love exploring possibilities to break out of these walls of texts that we all find ourselves in every day now Steve RuizBirgitta Böckeler - State of Play: AI Coding Assistants - AI Native DevCon June 2026Birgitta Böckeler - State of Play: AI Coding Assistants - AI Native DevCon June 2026
-
Birgitta Boeckeler hat dies repostetBirgitta Boeckeler hat dies repostetIt feels like the battle lines are drawn everywhere I look these days, between AI enthusiasts on the one hand, AI skeptics on the other, each wrestling with a real, scary, existential threat while downplaying the other. We aren't living in the same reality. The wins are real, the catastrophic downstream consequences are real, but they're being felt by two entirely different groups who are drifting further apart. Where there is no natural feedback loop, we must create one. Celebrate the wins, but surface the costs. Knit the story back together, so you can land bigger, braver wins with lower costs over time. Tell the whole story. It's a better story! And whenever you find yourselves arguing over the nature of reality, *stop*. This is not a philosophical debate, it's an engineering problem. Ground the discussion in something real. "What would it take for you to be willing to ship code without reading it?" is grounds for a more productive debate. We are all on the same side, and we need each other more than ever. https://lnkd.in/gjFVr7KqAI enthusiasts are in a race against time, AI skeptics are in a race against entropyAI enthusiasts are in a race against time, AI skeptics are in a race against entropy
-
Birgitta Boeckeler hat dies repostetHelping coding agents reason about software design Birgitta Boeckeler tested different methods for helping coding agents keep AI-generated code maintainable. Among the methods she compared, the Modularity skills came out as the standout finding. The point I love most: > Better interpretation of acceptable high-import-count “hubs” - Remember the “god classes” found by my previous coupling analysis? The modularity skills also noticed these, but in both cases nicely pointed out that they have a purpose in the context of this application. I assume that is either due to the good prompting in these skills, or due to the fact that this analysis actually read what was in the code, whereas I asked the other one to only rely on the coupling data. That's the essence of Balanced Coupling. Counting dependencies tells you the number of dependencies. It tells you nothing about the design. What matters is the kind of knowledge shared across component boundaries. A single reference sharing extraneous knowledge can cause orders of magnitude more cascading changes than a hundred dependencies through well-designed integration interfaces. Read Birgitta's piece on Martin Fowler blog: https://lnkd.in/dbs7vV_q
-
Birgitta Boeckeler hat dies geteiltMy post on my first impressions of spec-driven development in October 2025 [https://lnkd.in/dc4m77-e] is still getting a lot of traffic to this day, even though the tools I looked at back then either don't exist anymore or have evolved quite a bit. But the general challenge of how to build workflows with agents to get them to do what we want, and to let them help us figure out what we want in the first place, is still very much current. So I was glad when Laura Tacho suggested that we record a conversation about spec-driven development as of May 2026, now out on the Thoughtworks Technology Podcast: https://lnkd.in/d8ia4NAz Turns out, we ended up discussing points relevant to all of the questions I asked 8 months ago: - One workflow to fit all sizes? - Reviewing markdown over reviewing code? - How to effectively separate functional from technical spec? - Who is the target user? - Spec-anchored and spec-as-source: Are we learning from the past?
-
Birgitta Boeckeler hat dies repostetBirgitta Boeckeler hat dies repostetYesterday I was happy to present about TDD in the age of AI at Platmosphere. Thanks Giulio Roggero for the invitation! This is based on my work with Thoughtworks' clients, discussion with my colleagues at Thoughtworks, and other resources, chiefly Ivett Ördög's Approved Scenarios idea. Here is the slide deck. It represents my own opinions, not my employer's. https://lnkd.in/d7A9vv_e Gentle reminder that if you want to train with me, you have two opportunities: - June 23, in person at Software Quality forum - July 7-9, online, with Avanscoperta (both trainings are in Italian. Links in the comments)A Harness for Behaviour: how to get AI to generate code that does what we intend, or "TDD in the age of AI"A Harness for Behaviour: how to get AI to generate code that does what we intend, or "TDD in the age of AI"
-
Birgitta Boeckeler hat dies geteiltThird update to my article on "Maintainability sensors for coding agents": The test suite as a regression sensor. Assuming that the tests are testing the right things (!), how do we know they will fail when we break something? Are coverage metrics enough? (Spoiler alert: They are not) https://lnkd.in/deQRARJ4
-
Birgitta Boeckeler hat dies geteiltPart 2 of my article on maintainability sensors is out now, where I write about what I learned when trying to establish sensors about coupling and modularity. Spoiler alert: While my experiences in part 1 with linting as a computational sensor were quite good, for good feedback about cross-file concerns I had to go beyond static analysis and use LLM judgment as a sensor. https://lnkd.in/d_zfCvVz
-
Birgitta Boeckeler hat dies geteiltI recently wrote an article about harness engineering for coding agent users, where I described a mental model for expanding a coding agent harness: a system of guides and sensors that increase the probability of good agent outputs and enable self-correction before issues reach human eyes. As a follow-up, I have now written about my practical experience with using sensors regarding code maintainability. We will publish this in installments, this first one is about basic static code analysis, coming up next will be my learnings about sensors for code modularity, and then the usefulness of test suites as regression sensors. https://lnkd.in/dXRsg8FF
Birgitta Boeckeler hat einen Beitrag kommentiert
1 Woche
Klapprechnertrennungsangst
Birgitta Boeckeler hat einen Beitrag kommentiert
1 Monat
While CodeScene has of course been around for a while, we took the opportunity this time of "blipping" quite a few code and test quality related tools, as they are becoming so much more important in the age of AI coding. And you all have done great work navigating that space in a balanced and practical way, based on your deep data-grounded expertise of what code quality actually is
-
Birgitta Boeckeler hat darauf reagiertBirgitta Boeckeler hat darauf reagiertHonoured to be appointed to the DSIT College of Experts, lending my expertise to @SciTechgovuk along with 70 other specialists to help shape national science and tech policies and programmes. #Science #Innovation #Policy #UKScience #DSIT #CollegeofExperts #DSITCollegeofExperts
-
Birgitta Boeckeler hat darauf reagiertBirgitta Boeckeler hat darauf reagiertI'm really excited to announce I've started a new role at Adaptive! It might surprise some of you that this is a Product Manager role - it certainly surprised me, and was not where I thought my career was going. But I'll be managing Aeron Open Source, which is back to my roots in so many ways: 1) Fintech and high performance 2) Blogging about The Disruptor got me into DevRel in the first place 3) Open Source Software 4) So many LMAX folks are at Adaptive, it's like coming home There are also lots of new challenges. I see successful products as being the result of a loop: engineering -> marketing and sales -> product -> engineering. Developer Advocacy sits a little in all parts of this cycle, but the one I've got the least experience in is Product - how do we use everything we learnt from educating and speaking to users and prospects to create a better product? This role will (I hope!) help me succeed in this area. I'm super excited about this. I hope to reconnect with the communities around fintech and performance, and still work with the Java and extended software community.
-
Birgitta Boeckeler hat darauf reagiertBirgitta Boeckeler hat darauf reagiertGreat connecting with the #thoughtworks #platformengineering alumni at PlatformCon LND Live day. As Luca Galante and Kaspar Von Grünberg said in the welcome: feels like family. 🫶 Ajay Chankramath Rickey Zachary Rahib Amin
-
Birgitta Boeckeler hat darauf reagiertTraces, not transactions, are the composable units of modern observability. But traces are not the right method for *understanding* agents. One customer question might spin up a supervisor agent, which hands off to other agents, each of which makes multiple model calls, tool calls, API calls, and storage transactions before it delivers the answer. A trace captures one thread and stops there, leaving you copy pasting thread IDs between browser tabs as you try to reconstruct the path. It is genuinely hard to describe just how freaking cool this is; you're just going to have to see for yourself. Agent Timeline, now generally available. 🥰 https://lnkd.in/gNuFXqZvBirgitta Boeckeler hat darauf reagiertIf you're debugging AI agents by copying timestamps between browser tabs, traces are only giving you part of the picture. Agents spin up supervisors, hand off between workers, fire dozens of tool calls, and touch your whole backend before returning a single response. A trace captures one thread of that, but the rest disappears. The telemetry unit agents actually need is the agent conversation, which includes the full arc of an agentic workflow, from the first LLM call to the downstream system spans, all bound together by a conversation_id. It's the same shift in distributed tracing that we saw when we moved to microservices. Back then, following a single-process call stack wasn't enough. Now, following a single trace isn't either. Agent Timeline is how Honeycomb renders it. You open a conversation, see every agent running in parallel, and drill into any span to get the prompt, completion, model, tool name, and error — then connect straight to the backend root cause without leaving the page. It's GA now for every Honeycomb customer. Learn more in Dan Juengst's blog (link in comments). #honeycomb #observability #AI #agenttimeline
-
Birgitta Boeckeler hat darauf reagiertBirgitta Boeckeler hat darauf reagiertJust off stage at DEVOXX Poland where I debuted a new talk titled System Drift. Great fun as it always is here and I’m pleased with the way it went. Need to tighten it a bit but I think this has got legs. Thoughtworks Devoxx Poland #complexity
-
Birgitta Boeckeler gefällt dasBirgitta Boeckeler gefällt dasWARNING: This is LinkedIn. You will see posts about something folks are calling "Spec-Driven Development", and many of them - not all, but most - will be encouraging you to write detailed specifications for big chunks of system functionality before having an agent generate all that code and shipping it to their - almost always completely imaginary - users. They will tell you that the "old way" - where we specify a little slice of functionality, implement that vertical slice and ship software one slice at a time - is "dead". Do not be fooled. What they're advocating *is* the old way. We call it "Waterfall Software Development". You design a whole bunch of features. You write a whole bunch of code. You do a whole bunch of testing and review and fix a whole bunch of problems. Then you ship it in one "big bang". This is a very, very (very) risky way of doing things. It's putting all your chips on 29. Let's say you design, code, test and ship features A-Z - what happens if (and this is highly likely), feature A isn't really what the users need? And feature D only exists to make feature A work? And feature N is only needed to handle the output of feature D? You can think of software requirements as networks of interrelated hypotheses and assumptions - guesses, basically - and their design consequences. Not all software features are born equal. People don't sign up for online banking because they enjoy logging into things. The odds of getting it right first time are astronomically remote, and it's going to cost you a lot of time and money to find that out. If you're barking up the wrong tree, you want to know as early - and as cheaply - as possible before you build your entire treehouse on top of it. You need to test your guesses - no matter how educated they might seem - in the real world (or as close to reality as you can get). And you will *inevitably* need to learn and adapt to that feedback. That's where the real gold is - not in the design or in the plan, but in the feedback and in the learning. Don't listen to anyone encouraging you to bet the farm on a single hand of poker. It doesn't matter who's writing the code. It's the feedback loops doing the heavy lifting in creating value.
Ehrenamt
Sprachen
-
English
Verhandlungssicher
-
German
Muttersprache oder zweisprachig
-
French
Grundkenntnisse