This is a submission for the Google I/O Writing Challenge
🎬 The Scene
Google I/O 2026 dropped a wall of announcements in two hours.
🔥 Gemini 3.5 Flash
🤖 Antigravity 2.0
🛡️ Firebase AI Logic
🌐 WebMCP
🎨 Stitch
🧠 Jules
👁️ Gemini Omni
The keynote sugar rush was real.
Every recap I've read picks one announcement and explains it. That's useful. But it doesn't answer the question I actually had after the livestream ended:
🤔 Which of these can I use TODAY, in a real project, without it blowing up in my face?
So I spent the last 48 hours building with four of the newest tools from I/O 2026. Not demo projects. Not "hello world." Real integration attempts into actual workflows.
Here's what happened. 👇
🛠️ The Four Tools I Tested
I picked tools that cover different parts of the stack:
| # | Tool | What It Does |
|---|---|---|
| 1️⃣ | Antigravity CLI 1.0.2 | Successor to Gemini CLI — agent orchestration |
| 2️⃣ | Gemini 3.5 Flash | New default model via AI Studio API |
| 3️⃣ | Firebase AI Logic | Client-side AI inference with security |
| 4️⃣ | WebMCP | Protocol that makes web apps agent-readable |
I tried each one for a specific task. Not a tutorial. A real thing I'd actually ship. 🚀
1️⃣ Antigravity CLI: The 129 Skills Nobody's Talking About
Everyone's writing about Antigravity's multi-model routing (Gemini + Claude + GPT-OSS in one CLI). That's cool. 🆒
But the thing that actually changed how I work is /skills.
Antigravity ships with 129 built-in skills. Not autocomplete rules — actual agent behaviors. Things like:
- 🔍
agency-code-reviewer— reviews staged changes before commit - 🤖
agency-agentic-search-optimizer— audits whether AI agents can complete tasks on your site - 📖
agency-codebase-onboarding-engineer— helps new devs understand unfamiliar repos
🧪 The Test
I tested the skill creation workflow on a real React/TypeScript project. One prompt:
"Create a skill that enforces TypeScript strict mode violations before any PR merge"
⚡ What Antigravity Actually Did
Step 1: Read tsconfig.json and package.json → understood the stack ✅
Step 2: Scanned src/ for existing type patterns ✅
Step 3: Ran git status → understood current state ✅
Step 4: Proposed SKILL.md + checker script + pre-commit hook ✅
Step 5: Asked for approval, then built all three ✅
Step 6: Created mock violations, ran hook against itself, verified ✅
✅ The Good
One prompt. Zero config files written by hand. The pre-commit hook is active right now and will block the next TypeScript violation.
⚠️ The Bad
The skill lives globally in ~/.gemini/config/skills/, not in the project directory. That means it's available across ALL projects on this machine. Convenient until you have 60 skills conflicting with each other. 😬
❌ The Ugly
Gemini CLI (open source, 10K+ contributors) shuts down June 18. Antigravity is closed source. Google moved developer tooling into its monetization stack.
That's a tradeoff worth acknowledging. 🫠
🏆 Verdict
The skill system is genuinely powerful. The closed-source migration is genuinely concerning. Both are true.
⭐⭐⭐⭐ (4/5)
2️⃣ Gemini 3.5 Flash: Fast, Cheap, and Missing One Thing
I hit the Gemini API via AI Studio to power a content summarization feature. Straightforward task: feed it 3,000-word articles, get back structured summaries.
⚡ Speed
Sub-second responses for most inputs. Noticeably faster than Gemini 1.5 Pro for equivalent tasks.
Gemini 1.5 Pro: ~2.3s average
Gemini 3.5 Flash: ~0.8s average ← 3x faster 🚀
🎯 Quality
Good at extraction and summarization. Struggled with nuance — when I asked it to identify the "controversial take" in an opinion piece, it often defaulted to the most prominent claim rather than the most provocative one.
💰 Cost
This is where it gets interesting. Gemini 3.5 Flash is priced aggressively for high-volume use. If you're building a tool that processes thousands of documents daily, the economics are real. 📈
🚨 The Thing Nobody's Mentioning
Context window behavior. At 128K tokens, it technically handles long inputs. But I noticed quality degradation past ~60K tokens — the model started missing details buried in the middle of long documents.
This matches what other developers are reporting but nobody's writing about.
🏆 Verdict
Excellent for high-volume, structured extraction tasks. Don't trust it for nuanced analysis of long documents without a retrieval layer.
⭐⭐⭐⭐ (4/5)
3️⃣ Firebase AI Logic: The Security Model Is the Story
Firebase AI Logic lets you run Gemini inference directly from the client — your web app or mobile app talks to Google's API without a backend proxy.
The I/O keynote made this sound like magic. 🪄
The reality is more nuanced.
🛡️ What's Genuinely New: The 4-Layer Security Model
┌─────────────────────────────────┐
│ Layer 1: App Check │ ← Verifies requests from YOUR app
├─────────────────────────────────┤
│ Layer 2: Firestore Rules │ ← Controls who can call the model
├─────────────────────────────────┤
│ Layer 3: Rate Limiting │ ← Per-user throttling
├─────────────────────────────────┤
│ Layer 4: Output Filtering │ ← Content safety on responses
└─────────────────────────────────┘
This matters because client-side AI has always had a trust problem: if the API key is in the browser, anyone can abuse it. Firebase's approach doesn't eliminate that risk, but it adds enough friction that casual abuse becomes non-trivial. 🔒
🤷 What's NOT New
The inference itself. You could already call Gemini from a frontend using the AI Studio API. Firebase AI Logic wraps this in Firebase's auth and security ecosystem.
If you're already on Firebase → clean integration ✅
If you're not → migration cost is real ❌
🕵️ The Catch
Client-side inference means your prompt structure is visible in the browser's network tab. For any application where prompt engineering is part of your competitive advantage, you still want a backend proxy. 👀
🏆 Verdict
Great for Firebase-native apps that need AI features without backend complexity. Not a replacement for server-side inference in security-sensitive applications.
⭐⭐⭐ (3/5)
4️⃣ WebMCP: The Announcement That Could Matter Most (But Doesn't Yet)
WebMCP is a protocol that lets web applications expose structured information to AI agents. Think of it as robots.txt but for agent interactions — it tells AI crawlers what your app can do, not just what pages it has.
🤔 Why This Matters
The entire agentic stack (Gemini agents, Antigravity, Jules, etc.) needs to understand web applications to interact with them. WebMCP is Google's attempt at making that standardized.
😐 Why I'm NOT Excited Yet
I tried implementing WebMCP on a small web app and found:
- 📚 Documentation is sparse — the I/O session covered it in ~4 minutes
- 🔧 Tooling is minimal — no CLI scaffold, no validator, no testing framework
- 📉 Adoption is zero — no major frameworks support it yet
- ❓ It's a Google proposal, not a standard — W3C/IETF involvement is TBD
🏆 Verdict
Watch this space. Don't build on it yet.
⭐⭐ (2/5)
📊 The Final Scoreboard
| Tool | Score | Use It If... | Skip It If... |
|---|---|---|---|
| 🤖 Antigravity CLI | ⭐⭐⭐⭐ | You want agent-powered dev workflows | You need open-source tooling |
| ⚡ Gemini 3.5 Flash | ⭐⭐⭐⭐ | You're building high-volume AI features | You need nuanced long-doc analysis |
| 🛡️ Firebase AI Logic | ⭐⭐⭐ | You're already on Firebase | You need server-side prompt protection |
| 🌐 WebMCP | ⭐⭐ | You can afford to experiment | You need something that works today |
💡 The One Thing That Changed How I Think
The skill file. Hands down. 🏆
Before I/O 2026, my AI workflow was:
Open chat → Paste context → Get answer → Copy result
Open chat → Paste context → Get answer → Copy result
Open chat → Paste context → Get answer → Copy result
...forever 😩
The skill file inverts that:
Define behavior once (SKILL.md) → Agent executes autonomously → Forever ♾️
That's not a feature improvement. That's a different programming model.
The accessibility reviewer I built is now skill #130 on my machine. It lives at:
~/.gemini/config/skills/soilsense-accessibility-reviewer/SKILL.md
Every future Antigravity session can invoke it. One prompt created it. No orchestration code.
💬 The Gemini 3.5 Flash benchmarks will be obsolete in six months. A skill file that enforces your team's standards on every commit — that compounds.
🎯 What Would You Build?
I'm curious what others are finding. Have you tested any of these tools on real projects? What worked? What broke? 🤔
Especially interested in:
- 🐧 Anyone running Antigravity CLI on Linux (I tested on Windows)
- 🔥 Firebase AI Logic in production (not just demos)
- 🌐 WebMCP implementations in the wild
Drop your experience below! 👇
The best I/O coverage comes from people who actually built things, not people who watched keynotes. 📺➡️🔨
Thanks for reading! If this helped you decide which I/O tools to try, drop a ❤️ and share your own experience in the comments.
Top comments (0)