Derrick Egersdörfer

ETH Zurich stress-tested AGENTS.md files on Claude Code, Codex, and Qwen Code across 438 real tasks. LLM-generated context files made agents worse. Success rates dropped, costs jumped 20%+, reasoning tokens up 22%. The agents followed every instruction faithfully - that was the problem. More instructions meant more aimless exploration. Turns out agents are already good at navigating code. Most of what we're putting in these files is stuff they'd find on their own with grep. Human-written files helped slightly, but only when kept brutally short. The paper's advice: skip LLM-generated files entirely and keep human-written ones to minimal, essential requirements only.

Cursor shipped Composer 2 today - their own coding model, not a Claude or GPT wrapper. RL-trained inside real codebases. Full editor tool access. $0.50/M input tokens. The company that made "AI editor" a category just decided "editor" wasn't enough.

We keep making the same mistake. We watch the same five labs and assume that's where the next thing comes from. Meanwhile a phone company had the top model on OpenRouter for a week and nobody even thought to check.

Xiaomi hired ex-DeepSeek researchers, pointed them at their own GPU infrastructure, and shipped. Stock jumped 5.8% on the reveal - partly the model, partly their upcoming SU7 facelift. The model was already #1 before anyone knew who built it. That part was all performance.

For a week, the most popular model on OpenRouter was something nobody could identify. "Hunter Alpha." No branding. No launch event. Just burning through 500 billion tokens a week while the timeline argued over whether it was DeepSeek V4.

ChatGPT user time down 22% since mid-2025. Market share from 69% to 45%. A year ago a new model would own your feed for days. Now it gets a few hours before everyone moves on. Nobody's switching off AI. They just stopped caring which version number they're on.

Built a thing that lets Claude control a real browser. No Selenium scripts, no Playwright setup - just describe what to test. We're using it for integration tests. Claude navigates the page, does the thing, then we check the DB to confirm. Rust, open source. If you're into Claude Code or MCP tooling, come break it: github.com/codeChap/mcp-s…

Claude rate limits are doubled at night until March 28th. Anthropic's got spare capacity off-peak and they're letting you use it. Worth knowing if you run agents overnight.

@eddiejaoude No!

Do you watch your AI token usage?

The Claw ecosystem is out of control. Someone actually rewrote OpenClaw from scratch in Rust. 5 MB of RAM. Boots in under 10ms. Runs on a Raspberry Pi. They called it ZeroClaw. Seriously, how many Claws exist now?

Alibaba tested AI coding agents on real maintenance work - not toy benchmarks. 75% broke working code over 233 days. AI can write code. It can't maintain it. That distinction is your entire career.

@SamNewby_ @errgentai Nice!

@errgentai has tracked it's first 500 errors

Anthropic's new Cowork Dispatch: Text Claude a task from your phone. Go make lunch. Come back to your work on the desktop. Same conversation, same context, no re-explaining what you need. It's Mac-only, Max subscribers first, and your Mac has to stay awake - but the workflow already clicks. Early research preview.

Your SaaS product sells tools to humans. The next version sells agents that use those tools themselves. I avoided OpenClaw for months because it had zero security. Full network access, code execution, no guardrails. Couldn't put it anywhere near a corporate network. NVIDIA just wrapped it in something called NemoClaw. Sandboxed execution, policy engines, privacy routing. Agents locked down before they touch your data. Jensen compared this to Linux and HTML - the open standard that kicks off a whole era. He's underselling it. Every SaaS company that doesn't have an agent strategy in 12 months is already behind.

Jensen Huang closed GTC with singing robots around a campfire and Olaf from Frozen on stage. Every other tech CEO should be taking notes. #GTC

@henkjan Yip, it’s so important.

@CodeChap With AI, clarity matters more than speed.

A year ago the hard part was building it. Now I can describe something to Claude and have working code in minutes. The bottleneck is explaining what I actually want. Vague requirements used to mean slow progress. Now they mean sprinting in the wrong direction. The most dangerous dev on your team isn't the slow one. It's the one feeding AI unclear specs and shipping whatever comes back.

OpenAI's VP of Research, Max Schwarzer, just left for Anthropic. Said the people he trusts and respects most are already there. Their model policy lead went to Anthropic's alignment team. Their hardware lead quit over the Pentagon deal. A researcher dropped a resignation essay in the New York Times over the ad strategy. Anthropic refused the same military contracts OpenAI took. They're winning the majority of new enterprise matchups. "Benefit all of humanity" was always just the tagline.