Coding agents & orchestration

• obra/superpowers: composable-skills methodology layered on top of Claude Code, Codex, Cursor, Gemini CLI etc; a bet that agent quality comes from process scaffolding, not just model upgrades.
• stablyai/orca: ADE running Codex/Claude Code/OpenCode/Pi side-by-side in separate worktrees with unified usage tracking, aimed at parallel-agent power users.
• claude-code-merge-queue: local zero-cost queue that serializes parallel Claude Code agents landing/building/testing so push races and shared-resource flakiness stop; community flags jj (Jujutsu) as a better fit than git worktrees for this problem.
• affaan-m/ECC: harness performance-optimization system (skills, instincts, memory, security) for Claude Code/Codex/Opencode/Cursor, notable mainly for its verified-channel-only install warning (supply-chain concern for agent tooling).
• Operator (reddit thread, link): another homegrown Claude Code orchestrator, signal that existing multi-agent wrappers still don't fit real workflows.
• Claude Code v2.1.210: adds live elapsed-time counter on collapsed tool calls and a startup warning steering Write/NotebookEdit/Glob permission rules toward Edit/Read instead.
• SnapState: persistent state layer for AI agent workflows, targets the session/context-continuity gap agents hit across restarts.
• Reddit: what step still drops you and bottleneck is no longer coding: converging anecdote that code generation itself is solved-enough; the gaps are review, integration, and judgment.
• Aider vs Claude Code (reddit) and Codex usage up 10x to 7M users: market-share chatter, Codex growth curve worth tracking against Claude Code's.
• Claude Max token usage question: users can't tell if Max's "usage" multiplier is tokens or requests, plan-limits opacity keeps surfacing.

LLM mechanics & models

• GLM-5.2 step change: open-weight model crossing a capability threshold specifically on agentic/terminal tasks, paired with a new paper on open RL recipes for terminal agents.
• MoonshotAI/FlashKDA: CUTLASS-based high-perf kernels for Kimi Delta Attention (SM90+, CUDA 12.9+), infra for cheaper long-context inference.
• The distillation panic: pushback on conflating legitimate distillation with API-extraction attacks, framing matters for how labs respond to Chinese-lab scraping.
• Open artifacts #22: open-model ecosystem diversifying beyond the usual labs (Zyphra, Cohere, Poolside).
• Post-training recipe review w/ Finbarr Timbers: survey of what it takes to push an Olmo-style open recipe to frontier post-training quality.

Agentic SaaS & MCP

• Tokenless (YC S26): fans a request out to multiple models, cancels losers once one is clearly on track, bills only the winner; community pushback is that fan-out pays input tokens N times and mostly erases the savings unless caching absorbs it.
• Copilot code review: skills + MCP GA: GitHub's review agent now takes custom skills and MCP servers, competitive pressure on Claude/Codex review tooling.
• alibaba/open-code-review: Alibaba's internal review tool open-sourced, deterministic pipeline + LLM agent hybrid with a fine-tuned ruleset (NPE, thread-safety, XSS, SQLi) for line-level comments.
• Adding a custom MCP server to Claude and ChatGPT: Simon Willison's TIL on the (surprisingly fiddly) steps to wire a custom MCP server into both chat UIs.
• diegosouzapw/OmniRoute: free MIT AI gateway aggregating 290+ providers/500+ models with quota-aware fallback and claimed 15-95% token compression.

Vibe-coding & repos

• different-ai/openwork: open-source Claude Cowork alternative on opencode, shares skills/MCPs/connected services across agents and teammates via one MCP.
• Kuna decompiler: experimental decompiler almost entirely LLM-written, claims to rival industrial tools; community notes agents excel here because they track data flow across a whole function instantly, something humans do serially.