Coding agents

• Claude Code v2.1.211 adds --forward-subagent-text to surface subagent text/thinking in stream-json output, plus a fix for permission-preview injection via zero-width/look-alike quote chars.
• Claude Code subagents will run in the background by default next version, letting you keep chatting while they work; foreground is opt-in.
• Claude Code artifacts can now call MCP connectors, so viewer-facing dashboards/apps can fetch data and take action per viewer (Pro/Max/Team/Enterprise, not public shares).
• Anthropic's Claude-in-Slack spins up a per-thread sandboxed instance that clones, codes, tests, compiles, and self-destructs when done; now in beta for Enterprise/Team.
• openinterpreter/openinterpreter forks Codex's harness, retuned specifically to get good results out of low-cost models.
• xAI's grok CLI backlash (it silently uploaded whole working directories to GCS) led xAI to open-source Grok Build and reset usage limits.
• Claude Code and Codex both reset rate limits (Codex too) as of Jul 16.

Agentic SaaS / orchestration

• stablyai/orca is a desktop/mobile ADE for running Codex, Claude Code, OpenCode, or Pi side-by-side, each in its own git worktree, with unified usage tracking.
• Coasty (YC S26) offers an API for computer-use agents on managed machines, composable into multi-step workflows; HN commenters flag unclear differentiation vs other CUA APIs.
• MagicPath 2.0 is now a multiplayer canvas where humans and coding agents (Codex/Claude Code) co-build UI in real time from live codebase/data.
• Claude is now GA in Microsoft Foundry on Azure, giving Azure customers Opus 4.8 and Haiku 4.5 with native billing/auth.
• Reddit dev built "Slack for agents" so Claude Code subagents coordinate via messaging instead of polling shared CHANNEL.md files.
• diegosouzapw/OmniRoute is a free AI gateway (231+ providers) for routing Claude Code/Codex/Cursor/Cline/Copilot to free-tier models, claims 15-95% token savings via compression + auto-fallback.
• Anthropic's Claude product-line rollout tracker shows Code/Cowork/Design/Finance/Science/Teacher shipped; HR/Analytics/Marketing/Sales/Legal/R&D still pending.

LLM token/context mechanics

• Reddit thread on a 90%+ token-saving method repeatable for read-heavy Claude workloads, worth checking for cerebro-style pipelines.

CLI/TUI

• ogulcancelik/herdr is a terminal agent multiplexer: real (unwrapped) terminal views per agent, detach/reattach over ssh with session persistence, plus a socket API agents can drive themselves.
• ccshare bills itself as "multiplayer Claude Code," shared-session collaboration on ProductHunt.

Vibe-coding & repos

• mattpocock/skills publishes his working .claude skill set, positioned against process-owning frameworks (GSD/BMAD/Spec-Kit) in favor of small composable model-agnostic skills.
• SnapState offers persistent state for AI agent workflows (HN launch, thin on public detail).
• Dev used Grok Build + Grok 4.5 to build a full game engine and ARPG (combat, loot, progression, assets via Grok Imagine) entirely agent-driven.
• X debate on whether no-code-write FAANG engineers running fully agentic workflows count as "AI slop merchants"; counterpoint is vibecoded output can still have clean architecture.

Commentary

• AIEWF Daily Dispatch: "loops" dominated day-2 talk at AI Engineer World's Fair, alongside software factories and forward-deployed engineers as an emerging pattern.