Claude Tag: Multiplayer Async Agents in Slack

• Claude Tag launch — Anthropic's Slack bot gives Claude its own team identity, per-channel memory, and tool access; each @-tag spins up an isolated sandbox that clones repos, runs code, then discards itself.
• AINews deep-dive — covers the async-agent wave (Shopify, Stripe, Ramp) that Claude Tag joins; marks the shift from 1:1 chat to org-wide ambient AI.
• Proactive monitoring — Claude Tag can watch channels, auto-respond, and react with ✅/❌ emojis when threads close — no prompt needed, configured in natural language.
• Karpathy take — frames Claude Tag as "inline with all other human activity," the key being seamless tool/integration/compute/memory wiring, not the Slack surface itself.
• 65% of Anthropic product code — bcherny reports their internal Claude Tag already generates 65% of new product code; Claude Code runs under the hood.

Coding Agents & Models

• Opus 4.8 — 69.2% SWE Pro, 4× better self-error-correction, stronger instruction-following, plus dynamic workflows enabling multi-day tasks via hundreds of coordinated subagents.
• Claude Code v2.1.191 — adds /rewind to resume from before a /clear, fixes background agents resurrecting after stop — two long-standing pain points gone.
• Claude Code v2.1.187 — sandbox.credentials blocks sandboxed commands from reading credential files; org-configured model restrictions now enforce through the model picker and ANTHROPIC_MODEL.
• Fable 5 hints in v2.1.190 — string "You've used your Fable 5 usage for this week" added; "purchased separately" removed — signals Fable 5 bundling into subscriptions with weekly quota.
• Qwen-AgentWorld — language world model trained to simulate 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) natively; collection on HuggingFace, blog — environment modeling as a first-class training objective, not post-hoc RL.
• revfactory/harness — meta-skill that takes a domain description and generates a full Claude Code agent team plus the skills they use, selecting from six pre-defined team-architecture patterns.

LLM Mechanics & Token Economics

• LLM code style and token costs — structured JSON responses with schema constraints cut output tokens ~40% because the model stops over-explaining; useful signal for agentic pipelines where output cost dominates.

Agentic SaaS & Orchestration

• stablyai/orca — desktop/mobile ADE for running Codex, Claude Code, OpenCode, or Pi in parallel worktrees from one UI, with account hot-swap and usage tracking across providers.
• GitHub Copilot BYOK — Copilot app now supports bring-your-own-key for OpenAI, Azure OpenAI, Anthropic, and local models; lets teams route agent sessions to their own contracts.

Vibe-coding & Repos

• tashfeenahmed/freellmapi — OpenAI-compatible proxy aggregating 16 free-tier LLM providers (~1.7B tokens/month) with smart routing and automatic failover; personal experimentation only.
• Forward-Future/loop-library — public catalog of repeatable AI-agent workflow loops for engineering, eval, ops, content, and design; machine-browseable so agents can self-discover workflows.
• RubyLLM — unified Ruby framework across major AI providers for chatbots, RAG, and agents; community note: lacks observability hooks and retries wipe history, making debugging multi-call sequences painful.