Skills & agent tooling

Skill libraries for agents keep multiplying — worth knowing what's out there and diverging in quality: mattpocock/skills (curated, engineering-focused, not vibe-coding), dotnet/skills (.NET team's official skill set with an accuracy dashboard), alirezarezvani/claude-skills (354 skills across 13 tools, broad but unvetted), and tt-a1i/archify (agent skill for one-shot architecture diagrams, HTML export).

Claude Code internals & features

Reddit thread claims Claude Code subagents overpay ~14% on prompt cache due to a cache-key mismatch Anthropic hasn't fixed — worth watching if you spawn many subagents, cost impact unverified. Claude Code shipped Artifacts: interactive shareable pages (PR walkthroughs, dashboards) generated from a session, beta on Team/Enterprise.

Coding-agent tools & infra

codebase-memory-mcp indexes repos into a knowledge graph via tree-sitter (158 languages), sub-ms structural queries, claims 99% token reduction for agent context — plausible efficiency play if benchmarks hold. Mouse pitches precision file-editing beyond string-replace for agents, but HN commenters call the patent claims non-novel (basic ed-style editing) — credibility questionable. SnapState offers persistent state for agent workflows, thin signal, no detail yet.

Model landscape & post-training

Interconnects (Nathan Lambert) running a multi-part series worth the full read: open vs closed models are on different exponentials (economics of the gap), reading today's open-closed performance gap (why the single benchmark number misleads), bets on open models mid-2026 (predicts open stays in perpetual catch-up), the distillation panic (pushback on "distillation attack" framing for API extraction), and frontier post-training recipe review w/ Finbarr Timbers (RLHF/post-training state of play).

Agent architecture research

The Log Is the Agent (arXiv) proposes event-sourced reactive graphs as the core abstraction instead of conversation-loop-first design — state/logging as first-class, not bolted on; relevant if you're designing agent frameworks rather than just using them.

Codex reasoning-token anomaly

GitHub issue reports GPT-5.5 Codex reasoning tokens cluster at fixed boundaries (516, 1034, 1552) — looks like a throughput-driven token cap causing degraded performance on complex tasks; commenters say they've switched to Claude over this.

Agentic SaaS / vibe-coding

MagicPath 2.0 repositions as a multiplayer canvas for humans + coding agents (Codex, Claude Code) to co-build prototypes live. Termi Protocol — 3D live visualization of agents building, thin on substance. Open registry for Claude agent/companion templates, community-driven.