My favorite is the CSS bomb because it keeps finding surprising attack surfaces inside technology we treat as inert. Headlong and fx are close behind: one asks what happens when an agent never stops thinking, while the other strips the harness down until it can start almost instantly. I want to try both ideas in a few experiments.
Article
Gareth Heyes turns permitted email HTML and CSS into UI actions, indirect prompt injection, tracking, token exfiltration, and password theft. The variety matters: sanitizing scripts is nowhere near enough when hostile markup shares a trusted interface.
A React-focused comparison of PDFium and PDF.js across fidelity, zoom behavior, forms, bundle cost, text selection, search, and accessibility. Its sensible conclusion is to test both against your hardest real documents, not inherit an old benchmark.
Dev Agrawal uses an agent-produced migration to explain Solid 2.0 from a React developer's perspective. That concrete diff is more useful than another abstract framework comparison because it exposes where similar-looking components have different runtime semantics.
Ian Vanagas covers measurement, compact repository instructions, MCP tool costs, prompt caching, model selection, and durable memory. The useful warning is to track quality too; a cheaper agent that retries or fails is not an optimization.
Library&tools
Secretive keeps SSH private keys inside the Mac Secure Enclave, where they cannot be exported, while exposing them through an SSH agent. Version 3 adds post-quantum keys on macOS Tahoe and moves request parsing out of process.
jit moves credentials out of plaintext files and replaces them with decoys, then supplies usable values to developer tools when needed. It is an interesting defence against compromised packages, prompt-injected agents, and leaked dotfiles.
Fig reimplements React's Fiber and concurrent rendering ideas in a smaller TypeScript runtime, while preferring platform primitives such as AbortSignal and native events. Familiar component semantics with deliberately different lifecycle and data APIs make it a useful experiment.
Rust Glancer explores a lower-memory alternative to rust-analyzer by freezing analysis on save and storing its index on disk. Queries trade some speed for loading only the data they need, while common LSP features already work.
Ai
Atlassian's mcp-compressor replaces a server's full tool catalogue with a small proxy that discovers schemas on demand. Their reported 70–97% reduction targets a real waste: paying for thousands of tool-description tokens before useful work starts.
Vercel Labs' experimental coding-agent harness is a roughly 6 MB Zig binary with a shell-like interface, WebAssembly builds, and a claimed 10-microsecond cold start. Minimal prompts, tools, memory, and startup work are the product rather than afterthoughts.
Headlong keeps an agent thinking in a self-guided loop between external interactions, treating user input as another asynchronous event. A persistent inner monologue is a strange and compelling alternative to request-bound agents and scheduled checklists.
A study of repository-level instructions found lower median runtime and output-token use while maintaining comparable task completion. It is useful evidence that good context can save work, though association across benchmark runs is not a universal recipe.
Other
The old Windows pipes screensaver recreated entirely with HTML and CSS, powered by PolyCSS. It is playful, unnecessary, and exactly the sort of browser demo that makes a mature platform feel surprising again.