Coding agents & orchestration

• anthropics/claude-plugins-official: Anthropic's own curated plugin directory for Claude Code, ships with explicit "verify trust before install" warning since plugins can bundle arbitrary MCP servers/files.
• CopilotKit/OpenBot: open-source AI coworkers, each gets own browser/files/tools, every action pre-authorized and post-logged, bring-your-own AG-UI agent, real answer to "how much access do I give an agent."
• Claude Code Dynamic Workflows orchestrating Codex: user report on cross-model orchestration, Claude Code driving Codex as a subagent peer, not just reviewer.
• DietrichGebert/ponytail: skill forcing agent to write minimum code (lazy-senior-dev heuristic), measured 54% less code / 20% cheaper / 27% faster on real Claude Code sessions.
• coolplugz: another Claude orchestrator product, time-saving angle, no technical differentiation given yet.
• Diet Claude: usage-limit tracker/guard for Claude, targets the 5-hour-window pain point directly.
• GitHub Copilot app Customize tab GA: Copilot app now wires MCP/tool customization into a first-class UI tab, GitHub catching up to Claude Code's plugin ergonomics.
• Flare: graph-first IDE, visualizes codebase as interactive map for agentic coding sessions.
• Shubhamsaboo/awesome-llm-apps: 100+ open-source agent/RAG templates (Apache-2.0), cross-model (Claude/Gemini/GPT/DeepSeek/Qwen), useful as a scaffold-and-steal library.
• Built an AI memory extension with Claude Code, used it on itself for 150+ releases: dogfooded memory-management extension, relevant prior art for context/memory architecture.
• One prompt, 499 agents, 26M tokens, 5hr Max-20x limit burned in 1hr: concrete data point on real-world fan-out cost/throughput at scale.

LLM mechanics & infra

• Agentic Context Management (arXiv): argues agent failures are mostly context-management failures not reasoning failures, proposes treating memory/cost as lifecycle+architecture problem (schema validation, predictive fetching).
• OpenAI Jalapeño beats Nvidia Blackwell (SemiAnalysis): first deep benchmark of OpenAI's in-house inference chip, independent analysis says it beats Blackwell on inference; community take: entrenches compute oligopoly rather than democratizing inference.
• Jalapeño first results (OpenAI): OpenAI's own numbers, higher throughput + lower latency per watt, first custom silicon in production.
• The full stack behind abundant intelligence (OpenAI): Altman's framing of OpenAI's vertical integration thesis (datacenters→chips→models→products), context for the Jalapeño push.
• AINews: Muse Glimmer and Spark, open weights return: Meta/MSL re-entering open-weights race one year after "Personal Superintelligence" essay, Dreamer acquisition + Muse Code momentum.
• 5 useful things from new post-training textbook: Nathan Lambert's RLHF/post-training book (Manning) ships, distills lessons from training open models.
• How to evaluate LLMs before production (GitHub): GitHub's eval lessons from building secret-scanning with LLMs, practical pre-prod eval checklist.

CLI/TUI

• Claude Code v2.1.246: warns on wildcard-before-subcommand Bash allow rules (Bash(git * main) also matches inserted options, security-relevant), adds Auto-mode classifier tab to /permissions.

Vibe-coding & repos

• 600-card mobile roguelike deckbuilder, built mostly by Claude agents: game's theme is literally "working with AI agents," meta case study in agent-driven dev.
• TauricResearch/TradingAgents: multi-agent LLM trading framework, v0.3.1 adds look-ahead filtering, crash-safe graph routing, checkpoint resume, Sonnet 5/Fable 5 support.

Agentic SaaS

• Purchase API by Agentcard: one API call lets an agent buy anything online, agentic-commerce infra play.
• Agnost AI: catches agent failures evals miss, positions as production monitoring layer above standard eval suites.
• akta.pro: private company data/signals API built for agent consumption, B2B data-for-agents niche.