`Antigravity` Login Authentication Failure After Auto-Update

// related

• #0001

  #0001Agents & toolsGoogle AI Forum1 hour ago

  `Antigravity 2.0` Report: `Gemini 3.5 Flash` Looping on Complex Agent Work

  40radar

  AntigravityAI coding agent IDE — automates work with Google models

  A long verification task exposed repeated rereads after compaction. The practical takeaway: watch read/write ratios and split large source-heavy jobs.

  • —The trace shows 160 file reads vs 2 writes, an 80:1 read/write ratio. That is a strong loop smell for coding agents.
  • —At least 4 restart cycles repeated the same workflow, LaTeX files, task.md, and gap_analysis.md; compaction likely erased useful working detail.
  • —Only 4 command runs happened across steps 351-776. For verification tasks, execution count matters more than planner chatter.
  • —Large inputs plus heavy agent rules can push context compaction early. Smaller task chunks and explicit checkpoints reduce wasted runs.

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

  #0002Agents & toolsGoogle AI Forum2 hours ago

  `agy` arm64 build crashes on 39-bit VA Linux kernels

  40radar

  Google AntigravityCoding agent CLI — Google's `agy` command-line tool

  The failure happens before --version, inside bundled TCMalloc. ARM64 Android/LXC setups are a bad target until the binary handles 39-bit VA cleanly.

  • —Affected version: antigravity 1.23.2-1776332230 arm64; crash exits with code 134 before useful CLI output.
  • —Kernel uses CONFIG_ARM64_VA_BITS_39; bundled TCMalloc assumes a 48-bit userspace VA and tries high mmap addresses.
  • —LD_PRELOAD=jemalloc and TCMALLOC_ADDRESS_BITS=39 do not help because allocator symbols appear compiled into agy.
  • —Android-hosted LXC cannot fix this from Debian userspace. Treat agy on ARM containers as unsupported until Google ships a rebuilt binary.

  Source: discuss.ai.google.dev/t/bug-agy-crashes-on-arm64/145425Read original →
• #0003

  #0003Agents & toolsGitHub Changelog3 hours ago

  `Gemini 3.5 Flash` is now GA in `GitHub Copilot`

  80radar

  A faster, cheaper coding model option is landing in the IDE. Near-Pro quality at Flash-tier cost makes it worth testing for routine implementation loops.

  • —GitHub Copilot is rolling out Google’s latest Flash-tier model as a generally available option, not a preview-only experiment.
  • —Early testing claims near-Pro coding quality with Flash-tier speed and cost, useful for high-volume autocomplete and edit cycles.
  • —The practical play is model routing: keep premium models for hard design calls, use Gemini 3.5 Flash for repetitive coding throughput.

  Source: github.blog/changelog/2026-05-19-gemini-3-5-flash-is-genRead original →