Coding agents

• Claude Code subagents now run in background by default: keeps main session free to keep chatting while subagents grind, foreground still available by asking.
• Claude Everywhere / internal Claude Tag: Anthropic's internal tool runs Claude Code under the hood, already writing 65% of their product team's new code.
• Anthropic prompt library: official ready-to-use prompts for common coding tasks, cuts blank-page cold-start for Claude Code users.
• GitHub Copilot in Visual Studio, June update: usage visibility, MCP server trust layer, first C++ Copilot scenarios.
• GitHub Copilot for JetBrains BYOK expansion: broader model-provider flexibility across all Copilot tiers in JetBrains IDEs.

Agentic SaaS: Claude Tag

• Claude Tag launch: tag Claude into a Slack channel, gets own identity/memory, proactive multiplayer teammate not just a bot.
• Official Claude Tag announcement: joins Slack with scoped channel/tool access, delegate tasks while you work elsewhere.
• Sandboxed execution per thread: each tagged thread spins an isolated sandbox (clone, code, test, compile), destroyed after; per-channel memory/permissions.
• Proactive behavior: works without explicit prompting per instructions, behaves differently per channel using memory + data access.
• Slack monitoring pattern: Claude watches channels, answers questions, drafts PRs, reacts with emoji when threads resolve, just add @Claude and describe intent.
• Beta availability: point at channel, give task, let it work; in beta on Slack for Claude Enterprise/Team.
• MagicPath 2.0: multiplayer canvas where humans + Codex/Claude Code agents design/build together in real time.
• MagicPath native in Claude Code's in-app browser: now runs as extended design canvas inside Claude Code itself.

Vibe-coding & repos

• stablyai/orca: ADE for running Codex/Claude Code/OpenCode/Pi in parallel worktrees, one dashboard, usage/rate-limit tracking.
• Juggler: open-source GUI coding agent for hands-on control, visual workbench with inspectable tool calls and editable context; HN notes ACP plugin compat as the real switching blocker.
• Star Fleet Math: Mac app running 20 parallel Codex-based agentic harnesses ("starships") to attack open Erdős problems via Lean 4; HN flags the harness as unreproducible/proprietary.

LLM mechanics & open models

• GLM-5.2 assessed as step-change for open agents: capability threshold worth tracking against closed frontier.
• SGLang serving GLM5.2 NVFP4 at 500 TPS on 8xB300: new TopK-V2 kernel, 18-34% better single-user interactivity, 6-11% higher peak throughput on real agentic coding workloads.
• Open artifacts #22: Zyphra, Cohere, Poolside: ecosystem diversifying, more orgs shipping open models beyond the usual labs.
• Open vs closed models on different exponentials: whether users keep paying large margins for closed frontier models is the defining economic question of 2026.
• Frontier post-training recipe review w/ Finbarr Timbers: podcast on what an Olmo-style open recipe needs to reach frontier post-training quality.
• Some ideas for what comes next, May 2026: Gemini Flash 3.5, open-closed balance shifts, emerging power struggles in AI ecosystem.
• Claude Fable 5 and AI safety: Anthropic's silent model-manipulation-on-research-queries policy now uses a classifier like other safety domains, partial fix to a raised concern.
• Codex $200 plan limit changes: podcast covers Codex burning entire usage limit in one prompt, plus Windsurf/Codex-app churn.

CLI/TUI

• herdr: terminal agent multiplexer, real per-agent terminal views (blocked/working/done), detachable with SSH-survivable sessions, exposes a socket API agents can drive themselves.