Verdict: Google Antigravity 2.0 with Gemini 3.6 Flash is the most practical free agent setup for non-coders in 2026. The standalone desktop app (launched May 19, 2026 at Google I/O) replaced the original IDE-only Antigravity 1.0 with a five-surface platform — desktop app, CLI (agy), SDK, Managed Agents API, and Gemini Enterprise — built around dynamic subagents and scheduled tasks instead of a code editor. On July 21, 2026 Google made Gemini 3.6 Flash the default model, cutting output cost to $7.50 per 1M tokens (down from $9 on 3.5 Flash) while posting real benchmark jumps. The win is not the model alone — it is the combination of a subagent-capable harness with the cheap, fast workhorse model and a memory layer you actually feed.
Last verified: 2026-07-31 · Pricing, limits, and model versions change often — re-check before committing. Best free starting point: Antigravity 2.0 desktop app + Gemini 3.6 Flash default · Best for terminal-first:
agyCLI · Best for teams: Managed Agents in Gemini API
What is Google Antigravity 2.0 and how is it different from the original?
Antigravity 2.0 is Google's agent-first development platform — a standalone desktop app for macOS, Windows, and Linux that lets you launch, monitor, and orchestrate multiple autonomous agents in parallel. It is not a code editor. You talk to agents, watch what they produce, and give feedback on the work.
The November 2025 v1.0 was an agent-first IDE built on a VS Code fork. On May 19, 2026 Google recast the whole stack as a runtime for autonomous coding and knowledge-work agents, distributed across five surfaces that share one agent harness ([Antigravity product page, antigravity.google]):
| Surface | What it is | Best for |
|---|---|---|
| Desktop app (2.4.x) | Standalone agent command center, no editor | Knowledge workers, project oversight |
Antigravity CLI (agy, v1.1.x) |
Terminal agent runner sharing the same harness | Developers scripting agent tasks |
| Antigravity SDK (v0.1.x) | Python SDK over the harness | Programmatic, code-first automation |
| Managed Agents (Gemini API) | Google runs the harness; you call agents via API | Teams, server-side workloads |
| Gemini Enterprise Agent Platform | Managed enterprise deployment | Org-wide rollout, SSO/audit |
A clarification worth stating plainly: the Antigravity IDE is still available as a separate download on the same page (currently v2.1.x), for developers who want the editor surface. The "no code editor" framing applies to the new standalone app, not to the ecosystem as a whole — Google kept the IDE for people who write code.
What changed for agents in version 2.0?
Three architectural shifts make 2.0 the moment Antigravity stopped being a "smarter editor" and became a real multi-agent platform:
- Dynamic subagents. The main agent can now spawn its own subagents for focused subtasks, each with an isolated context window, running in parallel. You describe a goal in natural language; the orchestrator decides how to decompose it and creates the subagents it needs. This avoids the context-window degradation that sinks long single-agent runs. ([DataCamp: Antigravity CLI tutorial, July 2026])
- Background and scheduled tasks. Tasks can run in the background instead of blocking everything, and
/schedulelets you put an agent on a timer (one-off or cron-style) so it runs without you opening the app. ([Antigravity Getting Started docs]) - Projects (not repositories). Agents are now grouped into projects, and a project can cover several folders with its own settings and permissions — plus JSON hooks to intercept and control what the agent does, and live microphone transcription instead of shipping audio files to the model.
If your current agent workflow is one agent trying to do everything and burning out its context window, the parallel-subagent model is the upgrade that matters most. For the broader pattern of why single-agent setups bottleneck teams, see our piece on finding the bottleneck in your AI agent workflow.
Why does the Gemini 3.6 Flash default matter?
On July 21, 2026 Google dropped Gemini 3.6 Flash into Antigravity as the default model — quietly, with no code change required on the Managed Agents side ([Google Developers Blog: What's new in Managed Agents]). It also lives in Google AI Studio, Android Studio, the Gemini app, and Gemini Enterprise, so the skill travels across the surfaces you already use.
The numbers that actually matter for agent work, from the official [Gemini 3.6 Flash Model Card, Google DeepMind, July 2026]:
| Benchmark | Gemini 3.5 Flash | Gemini 3.6 Flash | Why it matters for agents |
|---|---|---|---|
| DeepSWE v1.1 (long-horizon coding) | 37% | 49% | Agents loop: try, check, fix, retry. Fewer loops = faster delivery. |
| OSWorld-Verified (computer use) | 78.4% | 83.0% | Click/type/UI automation accuracy — the foundation of browser agents. |
| Terminal-bench 2.1 (terminal coding) | 76.2% | 78.0% | Tool calls and shell commands land cleaner. |
| SWE-Bench Pro (agentic coding) | 55.1% | 58.7% | End-to-end multi-file coding tasks. |
| MLE-Bench (ML engineering) | 49.7% | 63.9% | Currently the best published MLE-Bench score of any tracked model. |
Pricing (verified, per 1M tokens): input $1.50, output $7.50, cached input $0.15 — versus 3.5 Flash's $1.50 / $9.00 ([BenchLM.ai Gemini 3.6 Flash profile, July 29 2026]; [DeepMind Model Card]). That is a 17% drop in output cost per token — and because agents are output-heavy (they write a lot), this is the line item that moves your bill.
Google's own framing labels 3.6 Flash its "workhorse model": better coding, better knowledge work, better multimodal (images and documents alongside text), and on the Artificial Analysis index it uses 17% fewer output tokens than 3.5 Flash to finish the same multi-step jobs. Both the per-token price and the token count drop at once — that compounding is the real reason to upgrade, not the headline benchmark.
For a deeper look at how grounding and verification change the math for agent-quality output, see our guide on why off-the-shelf AI gives bad financial advice (and the grounding fix) — the same dynamic applies to any agent producing load-bearing content.
How do you install Antigravity 2.0?
Install from the official download page: antigravity.google/download. The page detects your OS and offers the right build ([Antigravity Download page]).
macOS (minimum macOS 12 Monterey):
- Apple Silicon:
Antigravity.dmg(darwin-arm) - Intel:
Antigravity.dmg(darwin-x64) — note Apple no longer ships security updates for very old Intel macOS; Google's minimum is "the current and two previous versions."
Windows (minimum Windows 10, 64-bit):
- x64:
Antigravity-x64.exe - ARM64:
Antigravity-arm64.exe
Linux (minimum glibc ≥ 2.28, glibcxx ≥ 3.4.25 — e.g. Ubuntu 20.04, Debian 10, Fedora 36, RHEL 8):
- x64 or ARM64 tarball
For terminal-first work, install the agy CLI in 60 seconds:
# macOS / Linux
curl -fsSL https://antigravity.google/cli/install.sh | bash
# Windows PowerShell
irm https://antigravity.google/cli/install.ps1 | iex
If you already had the old Antigravity IDE: the next time the IDE auto-updates, it becomes "Antigravity 2.0" and prompts you to decide whether to keep the IDE alongside. If you write code, keep both apps — they sit in your dock with different icons. The standalone 2.0 app is agent-first; the IDE surface is still a VS Code-forked editor with an agent side-panel.
What are the slash commands every new user should learn first?
Per the [Antigravity Getting Started docs], four slash commands do most of the work. Learn them before you touch anything else:
| Command | What it does | When to use it |
|---|---|---|
/grill-me |
Forces the agent to ask you clarifying questions before it builds anything | At the start of any non-trivial task. Most bad output is really a bad brief. |
/goal |
Runs until the specified task is completely finished, asking no intermediate questions | Once the brief is locked and the plan is solid — not as your opener. |
/schedule |
Runs an instruction as a one-time or recurring timer (cron-style) | For anything repetitive: weekly reports, Monday content prep, daily health checks. |
/browser |
Explicitly directs the agent to use the browser subagent | When the task needs web navigation or visual UI testing. Agent judgment on "when to browse" is still imperfect, so this is explicit on purpose. /browser requires Google Chrome as your default browser to start a debugging session. |
Best-practice ordering: start with /grill-me, then /goal. Front-loading questions is the single highest-leverage habit — it converts a vague brief into a spec the agent can actually execute.
How do you write a multi-agent prompt that splits one job across roles?
This is the part most people skip, and it is the difference between "a chatbot that does five things badly" and "a team that ships." The technique is role-naming in the prompt: instead of asking for output, you ask for a team and let the orchestrator spawn the subagents.
Here is the general five-role pattern that works for any non-trivial multi-step deliverable (content series, proposal, research brief, onboarding flow):
Build a complete [deliverable] for [audience].
Requirements:
- [scope: e.g. five emails, friendly tone]
- [a constraint: e.g. introduce the community slowly, not all in message one]
- [a structural rule: e.g. end every item with a question]
Split the work between these roles:
1. Planner — decides the shape first (which item does what, how the story builds across the sequence)
2. Writer — drafts only once the plan exists
3. Reviewer — checks the draft against the original brief (did each requirement actually make it in?)
4. Critic — pushes on the writing itself: tone, flow, gaps
5. Auditor — catches what everyone else missed (e.g. an item silently ending without its required question)
What happens under the hood:
- The Planner subagent takes the requirements and decides the structure — which item introduces what, and how the narrative unfolds across the sequence instead of being dumped in item one.
- The Writer only writes once the map exists.
- The Reviewer checks the draft back against the original brief, so load-bearing constraints actually land in the output.
- The Critic pushes on the writing itself — tone, flow, clarity.
- The Auditor catches the quiet failures: an item that ends without its question, a missing call-to-action, a duplicated point.
You end up with a sequence that builds instead of five items that each try to say everything. Every prompt from that run lands in your memory layer, so the next time you build a similar deliverable the agents already know the voice.
This pattern generalizes far beyond content. The deeper principle — agents that only handle the gray zone, with clear handoffs between roles — is exactly what makes AI agents safe inside event-sourced systems and what lets teams simulate-test agents to ship faster before they ever reach production. The role-name trick simply makes that discipline explicit in the prompt itself.
What is the memory layer and why is it only as good as what you put in it?
Every Antigravity session stores its prompts and context in a shared memory layer. In practice this is often an Obsidian vault or similar markdown-backed store sitting underneath your agent stack, so all your agents — Antigravity, Claude, Codex, OpenCode, your own orchestrator — read from and write to the same place.
The point is simple: normally every agent starts from zero. You explain your brand voice once, then explain it again next week to a different tool. Shared memory kills that. You write your voice, your offers, and your formats down once and every agent in the stack works from the same page.
But the layer is only as good as what you put in it. If you skip the upfront work — dropping in your voice, your house style, your recurring formats — the memory is empty and the agents keep guessing. Treat it as deliberate infrastructure, not a bonus feature: a 30-minute investment once saves hours of re-explaining forever after.
What is the fastest realistic setup for a small team?
If you are a small business or solo operator who does not write code all day:
- Install the standalone desktop app from antigravity.google/download (keep the IDE out unless you actively code).
- Pick Gemini 3.6 Flash as your model — do not leave it on default if 3.5 Flash is set in any older session. On the Managed Agents API,
antigravity-preview-05-2026now defaults togemini-3.6-flashwith no code change required ([Google Developers Blog]). - Feed the memory layer on purpose — your brand voice, your offer stack, your standard formats, once.
- Front-load with
/grill-mefor anything non-trivial; switch to/goalonce the brief is locked. - Turn on
/schedulefor anything repetitive — weekly reports, Monday content prep, recurring audits. Anything on autopilot should be running on a timer, not waiting for you to remember. - Name roles in your prompt when the job is multi-step — the Planner/Writer/Reviewer/Critic/Auditor pattern above. Vague prompts get one generic agent; named roles get a team.
What this means for you
The upgrade that matters in 2026 is not a smarter model alone — it is the combination of a subagent-capable harness, the cheap fast workhorse model, and a memory layer you actually feed. Install Antigravity 2.0, pin Gemini 3.6 Flash, name the roles in your prompts, and schedule the repetitive work. Your first run will not go how you expect — you will write a too-vague prompt and watch one agent try to do all five jobs, or point a project at the wrong folders and wonder why the agent can't see your files. That is a setup problem, not a you problem, and it is fixable in one session with the steps above.
FAQ
Q: Is Google Antigravity 2.0 free? A: Partly. The desktop app, CLI, and SDK are free to download and use with your standard Google account. Bills start when you exceed your plan's quota on the Gemini API / Managed Agents tier — at $1.50 input and $7.50 output per 1M tokens for Gemini 3.6 Flash. At I/O 2026 Google offered AI Ultra subscribers $100 in bonus Antigravity credits (expired May 25, 2026); watch for similar offers.
Q: Do I need to write code to use Antigravity 2.0? A: No. The standalone 2.0 app has no code editor inside it — you talk to agents, watch output, give feedback. The separate Antigravity IDE (still available, v2.1.x) is the editor surface for developers. Most knowledge-work users never need the IDE.
Q: What is the difference between Antigravity 1.0 and 2.0?
A: 1.0 (November 2025) was a single agent-first IDE. 2.0 (May 19, 2026, I/O) is a five-surface platform — desktop app, agy CLI, SDK, Managed Agents API, and Gemini Enterprise — built around dynamic subagents, scheduled background tasks, voice input, and a default-model flip to Gemini 3.5/3.6 Flash. Google describes it as a full rebuild, not an incremental update.
Q: Which Gemini model should I use in Antigravity?
A: For most agent work, Gemini 3.6 Flash (the July 21, 2026 default). It posts DeepSWE 49% (up from 37%), OSWorld-Verified 83% (up from 78.4%), and costs $7.50/1M output tokens — cheaper than 3.5 Flash's $9. Use 3.5 Flash-Lite (gemini-3.5-flash-lite) for cost-sensitive, low-latency jobs where the quality bar is lower.
Q: What does /grill-me actually do? A: It forces the agent to ask you clarifying questions before it starts building. Use it at the start of any non-trivial task. It converts a vague brief into a spec — and most bad agent output is really a bad brief, not a bad model.
Q: Can I schedule agents to run without opening the app?
A: Yes. The /schedule command lets you run an instruction once at a future time or on a recurring cron-style schedule. Use it for weekly reports, daily content prep, and recurring audits so repetitive work runs on a timer instead of waiting for you.
Q: Where was Gemini 3.6 Flash released? A: July 21, 2026, by Google DeepMind — as the default for the Antigravity agent harness and available across Google AI Studio, Android Studio, the Gemini app, the Gemini Enterprise Agent Platform, and the Gemini API. The official Model Card and evaluation PDF are the primary sources for the benchmark numbers cited here.
Discussion
0 comments