Google Antigravity（AI IDE Platform Market Research）

1. Introduction

Antigravity is an agent-first IDE released by Google / Google DeepMind on 2025-11-18 alongside Gemini 3; at I/O 2026 on 2026-05-19 it was upgraded to version 2.0 and expanded into a full product family. It is Google's direct answer to the question "does the editor shape still matter?": it makes the Agent Manager (task console) — rather than the code editing area — the primary interface, shifting developers from "people who write code" to "people who review what agents produce."

Within this group, Antigravity's differentiation line is multi-agent parallel orchestration + reviewable artifacts (Artifacts): Agent Manager supports up to 5 agents advancing tasks in parallel, and each agent's plans, screenshots, and browser recordings are archived as Artifacts for review. It forms a succession relationship with the sibling Gemini CLI (see 13-gemini-cli.md) — after the Gemini CLI's personal free tier was discontinued on 2026-06-18, users were officially guided to migrate to Antigravity, which became the primary gateway for Google's developer agents.

1.1. Developer and Product History

1.2. Positioning and Product Family

The product family after Antigravity 2.0 covers five form factors:

Known pain point: the migration fragmentation caused by three apps coexisting (original IDE, 2.0 desktop App, CLI) is a concentrated source of community feedback — the official side has not published a unified full feature-parity table, [to verify].

1.3. Pricing

2. Glossary

3. Feature Description

3.1. Agent Orchestration Capabilities

3.2. Execution Surfaces and Extensions

The "editor + terminal + browser" three-surface integration is Antigravity's L2 signature: most competitors outsource browser verification to MCP tools or external automation, while Antigravity makes it a built-in execution environment — this is the fulfillment of the "agent-first" promise at the L2 level.

3.3. Enterprise Capabilities

4. Platform Architecture

Fig. 17-1 | Antigravity's Agent Manager orchestration and Artifacts review structure

Diagram: drawn based on this document's analysis.

5. Harness Design

5.1. L1 Context Engineering Layer

The distinguishing feature of L1 is the "experience dimension": most platforms' L1 focuses on what is seen in a single task, while Antigravity's knowledge/learnings library focuses on "what was learned in the last task." It runs in the same direction as Qoder's knowledge cards (see 14-qoder.md), but Antigravity binds it to the agent role (Custom Agent) rather than to the project.

5.2. L2 Tools and Execution Layer

Three-surface integration (editor + terminal + browser) + MCP structured tool execution preview. This is one of the most "full-stack" L2 execution-surface implementations in this group: the browser is no longer external but built in, and together with screen-recording capability it directly produces review evidence. Weakness: sandbox semantics are only visible in changelog fragments ("Inherit General"), and isolation primitives are not systematically disclosed, [to verify].

5.3. L3 Orchestration and Control Layer

Antigravity's L3 is one of the strongest orchestration layers in this group:

1. Parallel orchestration: Agent Manager runs up to 5 agents in parallel, monitored centrally from the console.
2. Chained workflows: one agent's output triggers the next agent, enabling composable pipelines.
3. Role-based customization: Custom Agents are declared via frontmatter and inherit skills, rules, and subagents.
4. Scheduling and remote: Scheduled Messages trigger on a schedule; Remote Control enables remote supervision.

Compared with Claude Code (see 02-claude-code.md): Claude Code's orchestration is "sub-agent dispatch within a single main loop," while Antigravity's is "console management of parallel multiple main loops" — the former is deep (isolation and coordination within a single task), the latter is broad (parallelism and supervision across multiple tasks). Choose the former for long-chain single tasks, and the latter for parallel multi-track work.

5.4. L4 Memory and State Layer

The L4 significance of Artifacts is consistent with this group's "state externalization" principle: it turns a task's intermediate state into a browsable artifact library, so people can review at any time "how far the agent has gotten" without relying on linear reading of the conversation history.

5.5. L5 Evaluation and Observability Layer

Artifacts are essentially an L5 design: it converts "done" into a verifiable chain of evidence — plan documents can be compared, screenshots visually inspected, browser recordings replayed, and test results re-checked. This is closer to the "reviewer's perspective" than a mere trajectory log. The enterprise tier provides audit trails. No public regression suite or automated scoring system on the product side, [to fill].

5.6. L6 Governance and Security Layer

L6 shows a layered structure of "strong enterprise tier, unclear personal tier": GEAP addresses enterprise governance needs, while the personal free/paid tiers' policies, audit, and data terms need to be checked item by item, [to verify].

5.7. Six-Layer Capability Summary

6. Case Studies

Note: As of the information cutoff date 2026-09-12, no customer quantitative case studies with controlled measurements published by Google were found; the following reflects public disclosures.

1. Migration succession case: On 2026-06-18 the Gemini CLI's personal free tier was discontinued)Skip, and users were officially guided to migrate to Antigravity; in 2026-03 Firebase Studio announced its deprecation on 2027-03-22 and guided users to migrate. These two migrations made Antigravity the sole primary gateway for Google's developer-agent line — this is a "case" at the product-strategy level, showing that Google has made the agent-first IDE the convergence point of its developer tools.
2. Changelog cadence: The continuous iteration from 2.9.1 (Remote Control) to 2.10.0 (embedded terminal + Git + audio attachments + structured tool execution preview) to 2.12.2 reflects a high-intensity engineering investment cadence (original page scrape failed, version details [to verify]).
3. Confirmable mechanistic conclusion: Agent Manager's "5-way parallel + Artifacts archival" combination is this group's most complete implementation of upgrading the "human-machine collaboration interface" from diff review to "artifact review"; this paradigm directly corresponds to the "verification cost is a real cost" thesis cited in this group's README (a lesson from the METR experiment) — the more reviewable the artifacts, the lower the walkthrough cost.

7. Summary

7.1. Strengths

1. Broadest orchestration: 5-way parallel + chained workflows + Custom Agents + scheduled triggers, with the strongest multi-task management capability in the group.
2. Artifacts paradigm: turns agent work processes into a reviewable evidence chain, directly addressing the "trust" gap.
3. Three-surface integration: editor + terminal + browser built in; frontend verification is not outsourced.
4. Open models: supports Claude, GPT-OSS, and more alongside the Gemini 3 series.
5. Google ecosystem integration: GEAP + Cloud integration, with a clear enterprise path.

7.2. Limitations

1. Migration fragmentation: three form factors — IDE, desktop App, CLI — coexist, and the full feature-difference table and migration path are not systematically disclosed.
2. Shallow sandbox and governance disclosure: isolation primitives, personal-tier governance granularity, and checkpoint mechanisms are all [to verify].
3. Unclear pricing: personal/team-tier pricing beyond the free preview is unconfirmed.
4. Strong Google ecosystem lock-in: migration succession (Gemini CLI shutdown) shows the risk of product lines shifting with strategy.
5. No public effectiveness measurement: customer quantitative case studies are missing.

7.3. Applicability Boundary and Selection Recommendations

The criterion for choosing Antigravity is: your working model is "advancing multiple task tracks simultaneously and accepting them one by one," and you can accept Google ecosystem lock-in. If single-task depth and sandbox governance matter most, choose Claude Code; if open source and customizability matter most, choose Codex CLI or OpenCode; Antigravity's irreplaceable differentiator is the "multi-agent parallel console experience."

Information Gap Statement

1. Personal/team-tier pricing structure beyond the free preview: not confirmed by an official page, [to fill].
2. Full feature-difference table and migration-path details between the 2.0 desktop App and the original IDE: not systematically sorted out, [to verify].
3. Sandbox semantics of Sandbox Mode (2.10.0 "Inherit General"): needs verification against official docs, [to verify].
4. Changelog version details (specific feature entries of 2.12.2, 2.10.0, 2.9.1): the official Changelog original page scrape failed, [to verify].
5. Context compression mechanism and checkpoint/rollback capabilities: undisclosed, [to fill].
6. Personal-tier data terms and governance granularity: undisclosed, [to verify].
7. Customer quantitative case studies: no official first-hand controlled data found; no marketing figures were adopted.

8. References

1. Antigravity official website — Google, 2025–2026. https://antigravity.google
2. Antigravity official Changelog — Google, 2026. https://antigravity.google/changelog
3. Antigravity encyclopedia entry (release information summary) — Baidu Baike, 2026. https://baike.baidu.com/item/Antigravity/67316199
4. Gemini CLI (sibling comparison; the party involved in the personal free tier shutdown event) — this group's 13-gemini-cli.md — Google, 2026.
5. Harnesses.sh (third-party Harness directory, including an Antigravity entry) — 2026. https://harnesses.sh
6. Introducing the Model Context Protocol — Anthropic, 2024-11-25. https://www.anthropic.com/news/model-context-protocol
7. Model Context Protocol official site — MCP / AAIF, 2024–2026. https://modelcontextprotocol.io/
8. Terminal-Bench official site (model + Harness system evaluation basis) — Stanford / Laude Institute, 2025–2026. https://www.tbench.ai/
9. Effective harnesses for long-running agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
10. Harness engineering: leveraging Codex in an agent-first world — OpenAI, 2026-02-11. https://openai.com/index/harness-engineering/
11. 2025 Stack Overflow Developer Survey — Stack Overflow, 2025-07-30. https://survey.stackoverflow.co/2025/
12. DORA 2025 State of AI-assisted Software Development — Google Cloud / DORA, 2025. https://dora.dev/