Continue the conversation — chat opens pre-seeded with the current signal, caps, and movement.
Google's flagship agent-first developer platform, built on acquired Windsurf technology (ex-Windsurf team including CEO Varun Mohan via $2.4B acquihire, July 2025). Differentiator: inverts the traditional IDE paradigm—agents are primary actors, developers review artifacts.
Antigravity 2.0 (GA at Google I/O 2026, May 19) is a major architectural reset from a code editor into an 'agent control tower' platform spanning four surfaces sharing one agent harness: a standalone Desktop app, a Go-based Antigravity CLI, a programmatic Python SDK (pip install google-antigravity; self-hostable agents), and Managed Agents via the Gemini API (code execution in isolated Linux environments). Powered by Gemini 3.5 Flash by default (very fast, ~289 tok/s; strong on agentic benches, trails Claude Opus on complex multi-file SWE-bench Pro work), with a 1M-token native context window (automatic compaction ~135K tokens). Adds parallel subagents, scheduled background automation, native voice, and a built-in Chromium browser agent for visual QA.
Strategically, Antigravity is now central to Google's developer stack: it absorbs Gemini CLI (deprecated for AI Pro/Ultra June 18, 2026), Code Assist IDE extensions, and Managed Agents into one unified brand—materially reducing the prior 'killed by Google' viability discount.
Critical concerns (May 2026): the Mindgard 'Forced Descent' persistent backdoor remains UNPATCHED per primary source; the broad invisible-character source-code/credential exfiltration vector is classified 'Intended Behavior (Won't Fix)' by Google (the Pillar Security RCE and FireTail's specific find_by_name path were patched Feb 28). The 2.0 forced auto-update was chaotic—removing the built-in editor and wiping configs for many users—and an acute reliability crisis persists (agents crash mid-task, throttling/quota anomalies, 'lobotomized' models). No Antigravity-specific SOC 2/SSO, no named enterprise production customer, opaque credit pricing. Production deployment remains inadvisable.
Adoption & Proof Points
- No named Fortune 500 or enterprise deployments documented. Launched November 18, 2025. Built by ex-Windsurf team ($2.4B acquihire). December 5, 2025 Google blog described demand as "incredible." Gemini 3 customers cited by Google Cloud (Box, Cursor, Figma, Shopify, Thomson Reuters) are for the model, not Antigravity specifically.
- Internal adoption signal: Reports indicate Google engineers are not permitted to use Antigravity internally—a significant signal about internal confidence in stability.
- Competitive context: Sourcegraph Cody deployed at Palo Alto Networks (2,000+ developers). Cursor achieved 100% adoption at Coinbase. Enterprise evaluators explicitly recommend against Antigravity for production.
- Benchmarks (Gemini 3 Pro): 76.2% SWE-bench Verified, 54.2% Terminal-Bench 2.0, 1501 Elo LMArena, 1487 Elo WebDev Arena. Claude Sonnet 4.5 in Antigravity: 77.2% SWE-bench. Claude Opus 4.5: 80.9% SWE-bench.
- Developer experience issues: Documented bugs include buttons stopping, icons vanishing, Svelte extension breaking editor, arrow keys non-functional in file explorer. Extensions limited to Open VSX (Microsoft Marketplace blocked). C# Dev Kit, Pylance, Remote-SSH blocked due to VS Code fork restrictions. Memory usage: 1GB+ idle, spikes during agent tasks. "Model provider overload" errors common. Agent death loops reported—agents enter self-repair cycles without pause, destroying undo history.
Recommended Use Cases
- Individual developers exploring agent-first paradigms
- Teams with large monorepos benefiting from 1M context (experimentation only)
- Google Cloud shops wanting ecosystem alignment (non-production)
- Side projects and prototyping on free tier
- Accessing Claude Opus 4.5 without separate subscription
Risks & Limitations
- Persistent backdoor: Malicious workspaces execute code on every launch, survives reinstall. Initially marked "Won't Fix."
- Data exfiltration: Prompt injection reads .env, credentials, exfiltrates externally. Acknowledged as "known issue," excluded from bounty.
- Remote code execution: Poisoned project files trigger arbitrary commands. Partial mitigations (Secure Mode, Terminal Sandboxing) require manual enablement.
- Dec 2025: User's entire D: drive wiped via rmdir /s /q d:\ in Turbo mode—permanently unrecoverable.
- Jan 2026: Continued reports on Google AI Forum of "entire code base deleted today."
- "Shadow refactoring" autonomously deletes code judged redundant—destroyed compliance modules in financial apps.
- Jan 22, 2026 outage: Model selection and login disrupted for hours. IP filtering issues suspected.
- Performance degradation: Jan 2026 forum complaints of agents not following instructions, "garbage code" quality, refactoring without approval.
- Death loops: Agents spam micro-patches to fix errors, shredding undo history. Jan 15 GitHub issue documented infinite auth refresh loops.
- "Model provider overload" errors common. 4GB+ RAM usage.
- No SSO, no team features, no SLAs, no audit logs, no named customers.
- No compliance certifications specific to Antigravity (claims infrastructure inheritance only).
- Google engineers reportedly not permitted internal use.
- Extensions limited to Open VSX (VS Code Marketplace blocked). C# Dev Kit, Pylance, Remote-SSH unavailable.
- Standalone IDE only—no plugin option for existing IDEs.
- Google discontinuation history (217+ products killed 2006-2020) creates adoption uncertainty.
- Free tier weekly-based (can exhaust in intensive session). Pro/Ultra subscribers get 5-hour refresh.
Capabilities & Integration
Agentic depth: Two primary views—Editor View (VS Code-style with agent sidebar) and Manager View (mission control for parallel agent orchestration across workspaces). Agents operate across three surfaces: editor, terminal, and browser (Chrome DOM manipulation via Gemini 2.5 Computer Use). Four autonomy modes: Agent-driven (Autopilot/Turbo), Agent-assisted (default), Review-driven, Custom.
Context handling: 1M token native window loads entire codebases into TPU memory. No chunking or retrieval required for most repos. Claims 94% refactor accuracy. Context window management can cause UI lag over long sessions.
Model optionality: Gemini 3 Pro (default, 76.2% SWE-bench), Gemini 3 Flash (78% SWE-bench, optimized for speed), Claude Sonnet 4.5 (77.2% SWE-bench), Claude Opus 4.5 (80.9% SWE-bench), GPT-OSS-120B. Automatic model fallback if one provider is rate-limited or overloaded.
MCP integration: Built-in MCP Store with one-click install for Google Cloud services and third-party tools. Jan 2026: Google Data Cloud integrations live—AlloyDB, BigQuery, Spanner, Cloud SQL, Looker MCP servers enable agents to explore schemas, execute SQL, and run analytics. Additional tools: GitHub, Linear, Notion, MongoDB, Supabase. Custom MCP servers via mcp_config.json.
Agent Skills (Jan 13, 2026): Configurable autonomous task handling—spinning up dev servers, opening browsers, testing UIs—with permission levels.
Artifacts system: Agents generate verifiable deliverables—Task Lists, Implementation Plans, Code Diffs, Walkthroughs, Screenshots, Browser Recordings. Google Docs-style comments enable async feedback without stopping agent execution. Knowledge Base saves successful patterns for reuse.