self.md radar — 2026-05-19
Agent work got dragged back to its boring dependencies: connector ownership, model economics, and whether the public record still has someone sweeping the floor.
The sharpest items today were not demos. Anthropic bought the API tooling layer, Cursor priced a coding model like margin pressure is the product, and a Papers With Code revival put maintenance back in the research loop.
1. Anthropic bought the SDK factory
sources:
what happened: Anthropic announced on May 18 that it is acquiring Stainless, the company that has generated every official Anthropic SDK since the early API days. Stainless turns API specs into SDKs, CLIs, and MCP servers across TypeScript, Python, Go, Java, Kotlin, and other stacks. Anthropic says hundreds of companies already use Stainless for the connectors that let developers, and agents, use an API.
why this matters: For agent products, connectors are not documentation. They are the hands. Owning the generator means owning one layer where API behavior, MCP shape, and developer defaults quietly harden into habit.
2. Cursor made coding models a price fight
sources:
what happened: Cursor released Composer 2.5 on May 18 and says it is better than Composer 2 at sustained work on long-running agentic tasks. The model is built on Moonshot’s Kimi K2.5 checkpoint, priced at $0.50/M input and $2.50/M output tokens, with a faster variant at $3/M input and $15/M output. Cursor also says it is training a larger model from scratch with SpaceXAI using 10x more total compute, after earlier Composer work hit a compute bottleneck.
why this matters: The coding-agent fight is moving into training stack, usage quotas, and compute access, not just editor chrome. A code editor now has to behave like a model lab, not merely a shell around someone else’s API.
3. Papers With Code came back as maintenance work
sources:
what happened: A Hugging Face open-source team member posted a revival of Papers With Code after the original site stopped being maintained after Meta acquired it. The new paperswithcode.co page is live with trending AI papers, code, datasets, methods, and evaluation leaderboards. The useful part is not a new benchmark. It is someone taking responsibility for the index researchers still use to find whether code, data, and comparisons exist.
why this matters: When benchmarks move faster than the index, evidence turns into private memory and scattered tabs. A maintained public map of papers, code, datasets, and leaderboards is dull infrastructure, which is exactly why losing it was annoying.
supporting links
- Sieve Secret Scanner — scans Claude Code, Cursor, VS Code Copilot, Windsurf, and Codex chat history for leaked secrets, all locally on Mac.
- SmallCode — local-model coding agent claiming 87% on a benchmark with a 4B-active model; good cost-pressure footnote, not another main coding-agent slot.
- Notebook Navigator 3.0 — Obsidian release adds manual sort, custom group headers, and word-count targets for people who still build work systems out of files.
- Handoff — tiny Show HN repo for preserving coding context when agents run out of tokens; thin, but it names a real workflow bruise.
left on the table
- llama.cpp MTP landed was a useful local-inference signal, but the underlying llama.cpp MTP PR already appeared in the May 17 seen ledger.
- Sieve stayed as an operator footnote because yesterday already covered chat leakage; this is the practical mop, not a new main story.
- Figure’s mail-sorting contest had reach, but the source was mostly spectacle and the operator takeaway was thin.
- Notebook Navigator 3.0 is useful, but it is a PKM release first and an AI-systems signal only by adjacency.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup