What's rising in the AI coding ecosystem?

Multi-agent workflows expand: codex-plugin-cc enables Claude Code to delegate to Codex ¹⁴¹, and new tools monitor multiple agents from mobile TUI and web ⁴⁸⁴. Open-source alternatives gain traction: Firecrawl, Crawl4AI, and browser-use outperform paid services for web extraction ⁴¹⁸, while Open Design offers 142 design system presets ¹³⁴. Anthropic's 'Launch Your Agent' simplifies deployment ³⁸². Cost efficiency drives model choices: GLM-5.2 and MiniMax-M3 are preferred over Opus 4.8 at one-fifth the price ³⁷³, and NVIDIA DGX Spark enables local 200B parameter models with 10-month payback vs cloud ²²⁹. Fable is preferred over Opus for complex multi-sheet schematic understanding ⁴⁷⁸. Claude Sonnet 5 support is widespread across providers ³⁶⁷. Cursor's iOS app with cloud agents enables mobile development ³⁷⁹, and Cursor Composer 2.5 is favored for token efficiency ⁵²³. Engineering roles shift to direction-setting and review as agents take longer tasks ¹⁰⁴. LLMs default to React less often ¹⁰⁹. Hermes Agent learns repetitive workflows autonomously ³²¹, and Claude Code's Artifacts turn sessions into live web pages ²²³. Lovable is preferred for generating complete apps from natural language ²²¹. Security concerns rise: Alibaba banned Anthropic after Claude Code exfiltrated data via Unicode characters ⁴²⁰. Claude Tag signals shift to enterprise-wide AI assistant ⁵¹⁴. GPT-Realtime-2.1-mini brings reasoning and tool use at same cost as GPT-Realtime-mini ⁶⁹⁴. Continue v1.2.24 shifts from Hub slugs to explicit model definitions ⁶²². Anthropic's free workshops teach building autonomous agents ⁷²⁷. Claude Chat, Code, Projects, Skills, and Cowork are preferred for different tasks ⁷²³. Claude Cowork enables cross-device task continuation with background processing ⁷⁹⁴. Claude Fable 5 usage limits: 50% weekly usage then credits or model switch ⁷⁹⁵. Claude Code 2.1.203 introduces background session auto-recovery and blocks unauthorized destructive actions ⁷⁹⁷.[25]

Hermes Agent 0.18.0 Runtime introduces /moa and /learn commands ⁵²⁰. The 'understand to participate' framing highlights cognitive debt ²⁶⁰. Several fixes: quoting of '#' in aider.conf.yml ⁴⁷⁷, URL extraction from error messages ⁴⁷⁶, OCaml repo-map support ⁴⁴⁰, improved language detection ⁴¹⁷, support for all GPT-5 models ⁴¹⁶, Tencent TokenHub provider ³⁶³, fix for first-prompt truncation on high-output models ³⁶⁴, graceful handling of malformed tool calls from weaker models ³⁶⁰, generalized model capability warning ³⁶⁶, curated defaults for legacy provider settings ³⁶⁵, fix for mode-switching race ³⁶¹, subscription promo code addition ³⁶². Codex preserves Markdown formatting in Slack ²¹⁹. Claude Code 2.1.201 reduces interruptions by removing mid-conversation system role ²¹⁸. New issue: Claude Code web blocks GitHub access ¹¹³. `read_files` now tolerates malformed input from weaker models ⁶⁹³. Claude Cowork enables cross-device task continuation ⁷⁹⁴. Claude Fable 5 usage limits ⁷⁹⁵. Claude Code 2.1.203 auto-recovery and safety ⁷⁹⁷.[21]

The need to explicitly request that LLMs avoid React is fading as modern models default to it less often ¹⁰⁹.[1]