Claude Tag (Slack agent)

Anthropic shipped Claude Tag: tag @Claude into a Slack channel, gets its own identity/memory per channel, works proactively without re-prompting.
Each tagged thread spins an isolated sandbox (clone, code, test, compile) that's thrown away after — one instance per thread, own permissions.
Anthropic's own product team reports 65% of new code now comes from their internal Claude-Code-based tool, "start of Claude Everywhere."
Setup is just adding @Claude to a channel and describing the behavior in natural language, it then monitors and reacts autonomously (drafts PRs, answers questions, reacts with ✅/❌).
Bernat notes it's proactive by design: no need to invoke it every time, it acts on standing instructions with per-channel memory.

Coding-agent adoption & economics

Peer-reviewed study of Microsoft's Claude Code + Copilot CLI rollout finds +24% PR merge rate, but community pushback notes velocity ≠ quality (no bug/incident data).
Codex usage up 10x in 6 months to 7M users, gaining ~1M/day, raising the "did Codex overtake Claude Code" question.
Reddit user priced all 12,992 requests in their Claude transcripts, single costliest message hit $7.72 for a one-line commit confirmation, good real-world token-cost data point.
Theo did a deep dive on Codex's "Ultra" mode, arguing it's misunderstood, worth reading before assuming it's just "more compute."

Agentic orchestration tooling

Orca: ADE for running Codex/Claude Code/OpenCode/Pi in parallel, each in its own git worktree, one unified view with usage tracking across providers.
MagicPath 2.0: multiplayer canvas where humans + Codex/Claude Code co-design and watch agents build prototypes live against real codebase/data.
SnapState: persistent state layer for AI agent workflows, aimed at surviving context loss across agent runs.

Context/harness engineering takes

Prediction thread argues agentic AI engineering collapses into prompt/context/harness/loop engineering converging back to plain software engineering + data-intensive agent design.
Matt Pocock flags "navigation pointers": thin link-chains acting as the roads agents use to traverse a codebase, a concrete pattern for making repos more agent-navigable.
system_prompts_leaks repo keeps extracted system prompts current across Claude Fable 5/Opus 4.8/Claude Code, GPT-5.6 family, Gemini 3.x, Grok, Cursor, Copilot, updated regularly, useful reference for prompt-engineering deltas.
Claude Fable 5 safety-fables writeup covers Anthropic's silent-classifier change on AI-research queries, relevant to understanding Fable 5's guardrail behavior.
OpenAI's Luna/Terra/Sol models land as new frontier models tuned for agentic/coding/scientific work with efficiency focus.

Vibe-coding & repos

page-agent: pure in-page JS GUI agent, no extension/headless browser needed, drives web UIs via text-based DOM manipulation (no screenshots).
Graphify: turns any folder (code, SQL, docs, media) into a queryable knowledge graph combining app code + schema + infra in one graph, works across Claude Code/Codex/Cursor/Gemini CLI.
DOOMQL: SQLite-as-game-engine Doom clone built with GPT-5.6 Sol, fun proof of "SQL owns movement/collision" concept.
Someone got Claude Code running natively on Windows XP (WIP, no SSH/RDP yet), niche but signals broad platform reach.
Fudge MCP: MCP server that feeds design taste from existing sites into agent workflows.
Jacquard: new lang for AI-written/human-reviewed code, explicit effects declarations for what functions touch globally, echoes Jai-style side-effect transparency.
Dad-built tool lets a whole team share one Claude Code seat instead of per-seat licensing.