`12-factor-agents`: production patterns for reliable LLM apps

// related

• #0001

  #0001Agents & toolsGoogle AI Forum12 hours ago

  `Google Antigravity 2.0` Cuts RAM, Breaks Migration Trust

  60radar

  Google AntigravityCoding agent IDE — agent-first development workflow

  The shift to an agent-first IDE removes direct editing and file-level warnings. RAM dropped from 1GB+ to 150-500MB, but migration risk makes it hard to adopt blindly.

  • —Forced updates moved users to Antigravity 2.0 without notice; side-by-side installs can overwrite the older IDE directory.
  • —Direct code editing and visual warning markers are gone. The workflow now favors agent commands over manual inspection.
  • —Runtime monitoring got more sensitive, but subagents misread graceful app shutdowns as failures. Process lifecycle still needs polish.
  • —Automatic Efficiency Mode cuts memory from 1GB+ to 150-500MB, making the tool viable as a background command center.

  Source: discuss.ai.google.dev/t/google-antigravity-2-0-review-beRead original →
• #0002

  #0002Agents & toolsGeekNews15 hours ago

  Microsoft Starts Pulling `Claude Code` Licenses

  40radar

  Claude CodeAI coding CLI — agentic code work from the terminal

  Internal tool access can flip when it threatens the platform owner’s stack. Keep coding-agent workflows portable; single-vendor lock-in is costly.

  • —Microsoft had opened Claude Code to thousands of developers since December, then began a broad license rollback after six months of heavy internal use.
  • —The cited pressure point is GitHub Copilot CLI: a rival coding CLI gaining traction inside Microsoft creates direct product conflict.
  • —Treat agent choice as replaceable infrastructure. Prompts, scripts, and repo workflows should survive a switch from Claude Code to another CLI.

  Source: news.hada.io/topic?id=29759Read original →
• #0003

  #0003Agents & toolsGoogle AI Forum17 hours ago

  `Google Antigravity 2.0` shifts to shared model quota

  60radar

  Google AntigravityAI coding agent IDE — supports multiple models per task

  Quota is pooled across Flash, Pro, GPT, and Claude instead of expanded. Better models get more usable headroom, but cheap prompts now drain the same weekly compute budget.

  • —The 5-hour reset can refill exhausted GPT and Claude pools to max, because quota now appears shared across models.
  • —Flash consumes less compute per prompt, while Pro or Claude burns more from the same pool. Minor edits should stay on cheaper models.
  • —If one model stops refilling, the whole weekly compute budget is likely low. Treat model choice as cost control, not just quality control.
  • —Google cites compute constraints and demand, so this is flexibility, not a quota increase.

  Source: discuss.ai.google.dev/t/the-new-antigravity-usage-model/Read original →