Obie Fernandez

We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord. Use this to message Claude Code directly from your phone.

Ticket-driven software development is slowly dying. For years, the workflow looked like this: - PM writes the ticket - Engineer picks it up - Engineer ships the code Repeat. That model made sense when writing code was the bottleneck. But it isn’t anymore. AI tools can generate large chunks of implementation. Scaffolding, tests, refactors, even entire features. The constraint moved. The hard part is now: - Choosing the right problem to tackle - Structuring the system correctly - Deciding what not to build That is why engineers are increasingly expected to own outcomes, not tickets. Instead of: - “Implement this API endpoint.” The work becomes: - What problem are we actually solving? - What’s the smallest thing we can ship that users will use? - How do we know it worked once it’s live? The engineers who adapt to this shift will thrive.

“LLMs are humanistic”

what's happening inside Anthropic right now is genuinely fascinating: DESIGNERS are now shipping production code with no engineer in the loop at all these are non-technical people with zero coding background. engineers on the team said designers are making "large state management changes you typically wouldn't see a designer making" think about what that means for a second. at every tech company in the world, designers and engineers are two separate roles. designers make mockups, engineers write code. that's been the deal for decades. at anthropic, that line is completely disappearing now. here's what their designers actually do now: it starts with a screenshot. 1) a designer takes their figma mockup and pastes the image directly into claude code. 2) claude looks at the image and generates a fully functional prototype from it. working code that engineers can immediately build on. this alone replaced the entire traditional cycle of: make static design → write spec → hand to engineer → wait → review → give feedback → wait again → repeat. but that's just prototyping. from there, designers implement front-end changes themselves. typefaces, colors, spacing, layout, etc directly in the codebase. no ticket filed, no engineer pulled in. and it goes further than visual polish. designers are now making state management changes (the code that controls how data moves through the app, even junior engineers find this tricky) and for their ongoing backlog of polish and bug fixes, they don't even open claude code. they just file a github issue (basically task description) describing what they want changed… and claude automatically proposes a code solution. then the designer reviews it and ships it. so there's a persistent stream of improvements flowing into production without pulling a single engineer off their work. they also use claude code to map out edge cases during the design phase. the stuff that used to only get discovered later when engineers actually built the thing. so now the designs are better AND they ship faster because the edge cases are already handled before engineering starts. one detail i love: they set up a custom memory file that tells claude: "you're working with a designer who has little coding experience. give detailed explanations. make smaller, incremental changes." claude adapted its entire approach based on that one file. here are some cool numbers, straight from the product design team: • claude code is now open alongside figma 80% of the time • 2-3x faster execution across the board. the designers' reaction when it all clicked: "holy crap, I'm a developer workflow" the boundaries between roles are dissolving. quietly. i think what's happening internally at anthropic is 2-3 years ahead of what will happen to the rest of the world

Tobi Lutke just pointed an autonomous AI researcher at the code that renders every storefront on Shopify. The agent found a 53% speedup. Liquid is the templating engine behind every single Shopify store. When a customer loads a product page, Liquid parses the template, executes the logic, and renders the HTML. That code path runs billions of times per day across 5.6 million active stores serving 875 million customers. A 53% reduction in combined parse+render time means every product page, every collection page, every checkout screen loads measurably faster. A 61% reduction in object allocations means less garbage collection, fewer memory spikes, lower compute costs per request. At Shopify’s scale, even single-digit improvements translate to millions in saved infrastructure. This is a double-digit overhaul. Four days ago, Tobi ran the same tool on a query-expansion model overnight. 37 experiments. 19% improvement. A 0.8B model outperforming the 1.6B model it was meant to replace. Now he’s running it against production infrastructure code that processes $292 billion in annual merchandise volume. The tool is Karpathy’s autoresearch: 630 lines of Python. An AI agent that modifies code, runs a training sprint, checks if the metric improved, and repeats. No human in the loop. Tobi pointed it at Liquid’s Ruby codebase and let it rip. 29 experiments run. 10 kept. 21 files changed. The screenshot shows the agent running benchmarks, discarding failures, and committing winners to a git branch. Tobi’s caveat that the results are “somewhat overfit” is the most important line. Benchmark numbers on a specific test suite rarely survive contact with production traffic patterns. But the ideas survive. The agent doesn’t just try random mutations. It reasons through the codebase, finds structural inefficiencies, and proposes targeted rewrites. The diff shows it replacing simple_lookup byte scan matching with regex, inlining method dispatches in the renderer, and swapping each/while loops for optimized for loops. The CEO of a $120 billion company is personally running AI research agents against his own core infrastructure on a Wednesday afternoon and posting the raw terminal output. That tells you more about where software engineering is heading than any product announcement this year.