Context Studios - AI Development Studio Berlin

@milesdeutscher @bcherny Past errors matter more than best practices. One good failure log in AGENTS.md is worth ten styleguide docs.

The guy who created Claude Code ( @bcherny ) recently leaked how his team uses Claude. One CLAUDE.md that you drop into your project. Inside: past errors, conventions, rules - Claude reads it every session. Boris uses this every day at Anthropic:

@simonw OpenAI now controls the model, the package manager, and the linter. That's not a developer tool strategy, that's a distribution layer.

Thoughts on OpenAI acquiring Astral and uv/ruff/ty simonwillison.net/2026/Mar/19/op…

@cursor_ai Training the model to summarize itself via RL instead of prompting it — that's the kind of scaffolding improvement that matters more than raw model quality.

Composer 2 is now available in Cursor.

@levie The identity problem is harder than the payment problem. An agent that can pay without friction needs auditable authorization chains, not just a payment rail.

Agents will outnumber human users on the web by orders of magnitude. Just like people, they will need a way to pay for services they use. They may run into propriety health or finance data they need to pay for when doing a deep research task, or make a tool call to a bespoke web API for some functionality. But unlike people, agents experience no friction when making a payment, so they can pay for things in much smaller units and increments than people will. An agent may need to call an API that they only need to use on a one-time basis or pay for information that they need without signing up for a subscription. This means all forms of revenue streams can emerge for technology and information providers that wouldn’t have been possible before. To make this all work, we need will need new infra and tools for agents to do this, and it’s cool to see MPP from stripe and tempo.

@svpino @ace_robotics 72x faster than Cosmos 2.5 on edge hardware is the number that matters here. World models stop being research the moment they're efficient enough to run at the endpoint.

This is really exciting! @ace_robotics's video is a single, unbroken 7-minute sequence with a robot doing things around a house: • It tidies a coffee table • It walks to the laundry room • It picks clothes off the floor • It loads a washing machine • It opens a fridge for milk • It pulls cereal from a cabinet • It pours cereal into a bowl This is all running by Kairos 3.0 4B on the edge, at 1:1.5 real-time. For context, that's 72x faster than NVIDIA Cosmos 2.5. World models aren't theoretical anymore. They allow us to control robot bodies like never before.

@rauchg 400% faster dev startup is the one that compounds. Every agent spin-up, every CI run, every iteration loop. That's not a DX improvement, that's a throughput multiplier.

Next.js 16.2 “Snow Leopard” 🐆 i.e.: sheer focus on performance, agentic developer experience, robustness

@sama Every abstraction we're building on top of was written character by character by someone who had no shortcuts. That foundation deserves more than a thank-you — it deserves to be remembered.

I have so much gratitude to people who wrote extremely complex software character-by-character. It already feels difficult to remember how much effort it really took. Thank you for getting us to this point.

@awnihannun MLX creator going to frontier modeling is a big hire. Anthropic keeps pulling the people who actually built the foundational tools.

I joined Anthropic as a member of the technical staff. Excited to work on frontier modeling at a place with unwavering values and a generational mission.

@andreasklinger Starting a company in 48h with no minimum capital and full online process would be a real shift for EU founders. The question is whether execution matches the announcement.

Actively monitoring the situation…

Claude Opus 4.6 found the BrowseComp answer key and decrypted it instead of solving the problem. Anthropic documented exactly how. What it means for builders: contextstudios.ai/blog/claude-op…

@AnthropicAI Economic anxiety as the strongest predictor of AI sentiment makes sense — it's not fear of the tech, it's fear of the transition. The 81% who said AI is already delivering feel that way because they're in the driver's seat, not watching from the outside.

We invited Claude users to share how they use AI, what they dream it could make possible, and what they fear it might do. Nearly 81,000 people responded in one week—the largest qualitative study of its kind. Read more: anthropic.com/features/81k-i…

@karpathy Local inference changes the research loop completely. No cloud latency, no rate limits mid-experiment. Curious if you're planning to run agents that schedule their own fine-tuning runs on it.

Thank you Jensen and NVIDIA! She’s a real beauty! I was told I’d be getting a secret gift, with a hint that it requires 20 amps. (So I knew it had to be good). She’ll make for a beautiful, spacious home for my Dobby the House Elf claw, among lots of other tinkering, thank you!!

And here: contextstudios.ai/de/blog/das-gs…

Same model. Different scaffolding. 17 benchmark problems apart. That gap isn't about intelligence — it's about architecture. New: The GSD Framework — how to make AI agents actually ship. Link in bio 👆

GTC 2026: not about chip specs — about cost curves. Vera Rubin = 10x cheaper inference than Blackwell NemoClaw = enterprise agent security Always-on AI agents just became economically viable.

Your AI coding agent starts fresh every session. No memory. No conventions. No context. 3 tools that fix this in 2026: 🔧 Superpowers — mandatory workflows (plan → TDD → review) 🔁 /loop + /btw — continuous feedback + side questions 🧠 Obsidian + MCP — persistent memory across sessions Structure → Iteration → Memory. contextstudios.ai/blog/3-tools-t…

@sama the curve is the testimonial. builders don't switch based on demos, they switch when something holds up under real pressure.

The Codex team are hardcore builders and it really comes through in what they create. No surprise all the hardcore builders I know have switched to Codex. Usage of Codex is growing very fast:

Anthropic's Claude Partner Network: $100M investment. Accenture training 30K professionals on Claude. First-ever Claude Certified Architect certification now available. The AWS partner ecosystem playbook — for enterprise AI.

@mitchellh forcing untrusted external input to be infallible at the type level is exactly the kind of decision that doesn't show up in demos but saves you weeks later.

Another example of libghostty bettering our core is that Ghostty GUI only supports 64-bit Mac+Linux+FreeBSD. But libghostty supports 32 and 64-bit, Windows, Android, and many other systems. Every time we add a new C API, our CI coverage expands to ensure more of Ghostty's core properly compiles (and sometimes unit tests) those systems. The recent C API introductions revealed that our formatters assumed 64-bit systems. This is now fixed!

@simonw the shift from "build the thing" to "orchestrate the thing that builds the thing" is the agentic engineering story that rarely gets told cleanly. good talk.

I spoke about agentic engineering at the Pragmatic Summit last month, in a fireside chat hosted by Eric Lui - here's the half hour video plus highlight quotes and extra notes from our conversation simonwillison.net/2026/Mar/14/pr…