Darragh Kennedy

AI productivity gains: More modest than expected newsletter.getdx.com/p/ai-productiv… < checks out, so far. But once the agentic operating model really takes hold, team structures change, and better platforms improve the build-to-prod stages, these numbers will jump.

This the load increase over 2 years that is breaking GitHub. Numbers on where the Y axis start provided by GitHub, added to the graph by me. Full analysis from today's @Pragmatic_Eng The Pulse issue

Token spend is out of control inside many (most?) tech companies. We gathered first-hand data from 15 companies on what is happening, and what they are doing about it. The two strategies are either "let it rip" or "curb spending, NOW" Full: blog.pragmaticengineer.com/the-pulse-toke…

Anthropic CISO just told you that 90% of their code is written by Claude. Then he explained how they protect their own secrets while doing it. Why your .env file is the weakest link in your entire AI workflow? Watch it, then grab the full security config below👇

MS: "All four are dramatically increasing capital expenditures driven by insatiable AI compute demand. Memory pricing inflation is a universal headwind, and management across all calls acknowledged they have consistently underestimated compute needs. Memory scarcity likely benefits cloud providers as suppliers prioritize largest customers." -META: CapEx guidance raised to $125–145B for 2026 (from $115–135B), driven by "higher component pricing this year and additional data center costs." Contractual commitments stepped up by $107B this quarter. "Our experience so far has been that we have continued to underestimate our compute needs." -GOOGL: CapEx guidance raised to $180–190B for 2026 (from $175–185B), including Wiz acquisition. Sundar Pichai noted: "We are seeing unprecedented internal and external demand for AI compute resources." Critically, 2027 CapEx expected to "significantly increase" vs. 2026. Cloud revenue "would have been higher if we were able to meet the demand"—i.e., compute-constrained. -AMZN: Q1 cash CapEx was $43.2B, "primarily relates to AWS and generative AI." Andy Jassy: "We expect Trainium will save us tens of billions of dollars of CapEx each year." AWS backlog stands at $364B (excluding the $100B+ Anthropic deal). Some investors are asking does that imply AMZN capex is actually decelerating given the big increase in memory costs and everybody else increasing capex on the back of memory. -MSFT: Calendar year 2026 CapEx expected at ~$190B, including ~$25B from higher component pricing. Q4 CapEx alone expected to exceed $40B. Amy Hood: "We expect to remain constrained at least through 2026." Added another gigawatt of capacity this quarter and on track to double overall footprint in two years.

@karpathy and I are back! At @sequoia AI Ascent 2026. And a lot has changed. Last year, he coined “vibe coding”. This year, he’s never felt more behind as a programmer. The big shift: vibe coding raised the floor. Agentic engineering raises the ceiling. We talk about what it means to build seriously in the agent era. Not just moving faster. Building new things, with new tools, while preserving the parts that still require human taste, judgment, and understanding.

I spent the last couple of weeks diving deep into Playwright tests and trying to understand and fix tests with zero domain knowledge. A task that would have taken way to long and way to much headache was possible thanks to AI and knowledge of the framework. Understanding how to use AI to help you, to analyse and learn and discuss improvements is key. Taking the time to do it is important. Many of us have so many tasks to do that we forget to take time to step back and just think for a while and research and then act. For a few days it looks like we are doing nothing as we have nothing to show but then...... a massive refactor that improves a whole codebases test. Now that is well worth stopping and taking time for. I wrote a post on the process and what and how I did it. Hope it helps someone. dev.to/debs_obrien/ho…

Ghostty is leaving GitHub. I'm GitHub user 1299, joined Feb 2008. I've visited GitHub almost every single day for over 18 years. It's never been a question for me where I'd put my projects: always GitHub. I'm super sad to say this, but its time to go. mitchellh.com/writing/ghostt…

Google is turning consultants into its AI delivery network with a $ 750M fund for firms like McKinsey, Accenture, and Deloitte to help companies build and scale agentic AI. Consulting firms need this because classic consulting work, such as research, slide drafting, process mapping, and software planning, is exactly the kind of work AI systems are starting to automate. AI startups need consultants because big companies rarely buy new tools just because the model is powerful, since they need someone to connect it with data, workflows, security rules, and staff habits. Agentic AI means software that does not only answer questions, but can plan steps, call tools, move through business systems, and complete tasks with less human steering. So Google’s bet is that McKinsey can find the business problem, Google can provide the AI stack, and the client can turn a pilot into a working system across teams. OpenAI’s reported push to sell Codex through Accenture, Capgemini, and PwC points to the same shift, where AI coding tools become enterprise software only after consultants package them into training, governance, and rollout plans. --- businessinsider. com/consulting-mckinsey-accenture-bcg-ai-silicon-valley-enterprise-partnerships-2026-4

Guess we’re about to find out how much the “endless demand” for AI narrative falls apart when the subsidy pricing goes away A 25x pricing increase probs nothing to be concerned with tho

Being the foundation for millions of developers means our bar must be higher for availability, reliability, and security. I’m sorry it’s been a rocky stretch at GitHub. We know we need to do better. Today we published an update on two recent incidents: one on April 23 involving merge queue behavior, and one on April 27 affecting pull requests, issues, projects, and search-backed experiences. We’re taking this seriously. We’re listening, and you have my commitment that we’ll communicate more frequently about the work underway to improve reliability and scale GitHub for what comes next. github.blog/news-insights/…

AI leaders need to talk more about the incredible positive upside of AI, not the conflated downsides (yes, I'm looking at you Dario). $GOOGL $NVDA THANK YOU to Google's Amin Vahdat and Jeff Dean for their inspiring visions on what AI will enable. The video clip below is a MUST WATCH. Also: “No matter what, it’s going to take a lot of compute” It would be the dumbest thing ever to overregulate and cripple AI progress based on false, misleading AI doomer narratives.

The Son of Anton scene will almost certainly be the most referenced Silicon Valley scene in age of AI: “It’s possible that…the most efficient way to get rid of all the bugs, was to get rid of all the software.”

@MTSlive @levie Thanks!

@darraghke @levie Coming soon on Youtube

.@levie on why enterprises need people to implement the agents that are supposed to automate the people: "If agents just get the exact same permissions that you had, then they'll just run into these walls everywhere." "Unlike a human, they're not gonna know to go talk to Sally or ask the question of Bob. So they're gonna just be kind of stuck." "A large enterprise is gonna have to go through the change management, the systems implementation, the integration of technology for these agents to be able to go and work." "People thought it was somewhat ironic that, oh, we need people to implement the agents that are gonna go automate the people, and it's like, no, that's exactly how it works."

Engineers at OpenAI experience the same problem as everyone else — we can supervise about 3–5 coding agents. After that productivity drops. Codex is smart, but our attention is limited. So we built (and open sourced!) Symphony to remove that ceiling. Here’s how it works:

Its the beginning of the end of subsidized AI subscriptions. GH Copilot is moving to usage-based billing, as has Claude (for business customers.) Fair to assume more will follow. I expect this change will also be a great boost for open models - cheaper, and pretty good already

𝗧𝗵𝗼𝘂𝗴𝗵𝘁𝗪𝗼𝗿𝗸𝘀 𝗧𝗲𝗰𝗵𝗻𝗼𝗹𝗼𝗴𝘆 𝗥𝗮𝗱𝗮𝗿 - 𝗩𝗼𝗹𝘂𝗺𝗲 𝟯𝟰 (𝗔𝗽𝗿𝗶𝗹 𝟮𝟬𝟮𝟲) Two big shifts in this edition: Claude Code moved to Adopt, and MCP by default moved to Caution. Both in the same Radar. The rest of Volume 34 is about agents, with warnings sharper than V33. Four themes stand out: 𝟭. 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗻𝗴 𝘁𝗲𝗰𝗵 𝗶𝘀 𝗴𝗲𝘁𝘁𝗶𝗻𝗴 𝗵𝗮𝗿𝗱𝗲𝗿 Semantic diffusion is part of it. Terms like spec-driven development and harness engineering get used inconsistently, often before their meanings settle. The pace makes it worse. Some tools ThoughtWorks reviewed were less than a month old and were often maintained by a single contributor using a coding agent. The bigger risk is 𝗰𝗼𝗱𝗲𝗯𝗮𝘀𝗲 𝗰𝗼𝗴𝗻𝗶𝘁𝗶𝘃𝗲 𝗱𝗲𝗯𝘁: as more code is generated by AI, teams adopt solutions without building the mental models to debug them later. 𝟮. 𝗥𝗲𝘁𝗮𝗶𝗻𝗶𝗻𝗴 𝗽𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲𝘀, 𝗿𝗲𝗹𝗶𝗻𝗾𝘂𝗶𝘀𝗵𝗶𝗻𝗴 𝗽𝗮𝘁𝘁𝗲𝗿𝗻𝘀 AI is pushing teams back to the foundations: 𝗽𝗮𝗶𝗿 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴, 𝗺𝘂𝘁𝗮𝘁𝗶𝗼𝗻 𝘁𝗲𝘀𝘁𝗶𝗻𝗴, 𝗗𝗢𝗥𝗔 𝗺𝗲𝘁𝗿𝗶𝗰𝘀, 𝘇𝗲𝗿𝗼 𝘁𝗿𝘂𝘀𝘁 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲. After years of being abstracted away in the name of usability, the terminal is back as a primary interface. But not everything carries over. Team topologies will need to evolve into agent topologies, and feedback cycles need rethinking. 𝟯. 𝗦𝗲𝗰𝘂𝗿𝗶𝗻𝗴 𝗽𝗲𝗿𝗺𝗶𝘀𝘀𝗶𝗼𝗻-𝗵𝘂𝗻𝗴𝗿𝘆 𝗮𝗴𝗲𝗻𝘁𝘀 The agents worth building are the ones that need access to everything. OpenClaw and Claude Cowork supervise real work tasks. Gas Town coordinates agent swarms across entire codebases. Simon Willison's "𝗹𝗲𝘁𝗵𝗮𝗹 𝘁𝗿𝗶𝗳𝗲𝗰𝘁𝗮" describes an unsafe agent: private data, untrusted content, external action. That now describes most useful agents by default, not by misconfiguration. Zero trust, least privilege, defense in depth. All table stakes. Safe agent systems will be pipelines of constrained agents, not monolithic ones. 𝟰. 𝗣𝘂𝘁𝘁𝗶𝗻𝗴 𝗰𝗼𝗱𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁𝘀 𝗼𝗻 𝗮 𝗹𝗲𝗮𝘀𝗵 As coding agents get better, humans are tempted to step out of the loop. Teams are pushing back with 𝗰𝗼𝗱𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 𝗵𝗮𝗿𝗻𝗲𝘀𝘀𝗲𝘀. Feedforward controls, such as Agent Skills, modularize instructions and load them just in time. Spec-driven frameworks like GitHub Spec-Kit and OpenSpec structure planning, design, and implementation. Feedback controls wire deterministic quality gates into agent workflows: compilers, linters, type checkers, mutation tests. Failures trigger auto-correction before human review. In more detail: 𝗧𝗲𝗰𝗵𝗻𝗶𝗾𝘂𝗲𝘀: ✅ Adopt: Context engineering, Curated shared instructions, DORA metrics, Passkeys, Zero trust architecture 🧪 Trial: Agent Skills, Feedback sensors for coding agents, Mutation testing, Progressive context disclosure, Sandboxed execution for coding agents 🛑 Caution: Agent instruction bloat, Codebase cognitive debt, Coding agent swarms, MCP by default, Pixel-streamed development environments 𝗣𝗹𝗮𝘁𝗳𝗼𝗿𝗺𝘀: 🧪 Trial: AG-UI Protocol, AWS Bedrock AgentCore, Graphiti, Langfuse, Replit, SigNoz 🔍 Assess: ClickStack, Coder, MCP Apps, Sprites 𝗧𝗼𝗼𝗹𝘀: ✅ Adopt: Axe-core, Claude Code, Cursor, Kafbat UI, mise 🧪 Trial: cargo-mutants, Claude Code plugin marketplace, Dev Containers, Figma Make, OpenAI Codex 🛑 Caution: OpenClaw 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲𝘀 𝗮𝗻𝗱 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸𝘀: ✅ Adopt: Apache Iceberg, Declarative Automation Bundles, React JS, React Native, Svelte, Typer 🧪 Trial: Agent Development Kit, DeepEval, Docling, LangExtract, LangGraph, LiteLLM

The only people who believe any of this are non-coders. I tried to build a game (an area I’m an n00b in.) The results are amusingly disastrous - I never before coded a decent game. But I’ll crack out backend services w AI rapidly - because I coded dozens of them before…