Google Antigravity + Felo Search: Give Your AI Agents Live Research and Cited Answers
Close Gemini 3's knowledge cutoff gap with Felo Search skill for Google Antigravity — live web search, cited answers, and real-time verification, auto-triggered inside the Agent Manager.
Google Antigravity's Agent Manager is a capable planner. It coordinates multi-step tasks, breaks down complex requests, and delegates work across sub-agents. But it has a fundamental limitation: every agent starts with only what Gemini 3 knows. And what Gemini 3 knows is frozen at its training cutoff.
That gap between planning capability and information access is where real work stalls. An agent can plan a competitive analysis, but it can't verify current pricing. It can outline a market report, but it can't check whether a product launched last week. It can draft a technical memo, but it can't confirm whether an API endpoint changed.
Felo Search bridges that gap. It's a live research skill that plugs directly into Antigravity's .agent/skills/ directory — no configuration, no slash commands, no manual invocation. When an agent task requires current information, Felo Search activates automatically and returns a cited answer the agent can act on.
This article explains how it works, what it unlocks, and why it should be the first skill most Antigravity teams install.
The Knowledge Cutoff Problem
Every language model has a training cutoff — a date after which it simply doesn't know what happened. For Gemini 3, that means:
- Stale pricing: Product prices change monthly, sometimes weekly. An agent working from training data will quote outdated numbers.
- Missing releases: New products, API versions, and feature launches happen every day. The model has no way to know about them.
- Outdated documentation: API specs deprecate, endpoints move, authentication patterns change. Training data is a snapshot; reality is a stream.
- Competitor blind spots: New competitors emerge, existing ones pivot. Agents need to know about moves that happened yesterday, not six months ago.
These aren't edge cases. They're the everyday information gaps that make an otherwise capable agent produce wrong answers, incomplete analyses, and outdated recommendations.
The solution isn't a bigger model. It's a live data layer.
How Felo Search Works Inside Antigravity
Felo Search integrates with Google Antigravity through a simple folder install. Drop the felo-search skill folder into your project's .agent/skills/ directory, commit to Git, and every developer's Antigravity instance picks it up on next pull.
Automatic Activation — No Slash Commands
The SKILL.md file in each Felo skill contains a semantic description of what the skill does. Antigravity's Agent Manager reads these descriptions and matches them against task requirements. When a task involves current pricing, release notes, market research, or any question where the model's training data isn't enough, the Agent Manager loads Felo Search automatically.
You don't type /felo-search. You don't configure triggers. You just ask the agent to do its job, and it reaches for live data when it needs it.
The Request Flow
Here's what happens under the hood when an agent task triggers Felo Search:
- Task analysis: The Agent Manager identifies that the current task requires information that may be outside Gemini 3's training window — current pricing, recent news, updated documentation.
- Skill activation: Felo Search is loaded automatically based on SKILL.md semantic matching.
- Query optimization: The skill reformulates the agent's question into optimized search queries.
- Live search execution: Queries are sent to Felo's real-time search API, which aggregates results from the current web.
- Cited answer generation: Results are synthesized into a structured, cited answer that the agent can use directly in its workflow.
- Task continuation: The Agent Manager continues execution with verified, current information — no context switching, no manual intervention.
The entire flow happens inside the IDE. The agent never leaves the conversation. The developer never leaves their workflow.
Structured, Cited Answers
Felo Search doesn't return a wall of text. It returns a structured response with:
- Comprehensive answer: A synthesized response that directly addresses the agent's question.
- Query analysis: The optimized search terms that were used, providing transparency into the research process.
- Source attribution: Results are grounded in current web data, not model hallucination.
This structure matters because agents need reliable inputs. A cited answer lets the agent verify its own reasoning, trace back to sources, and produce output that a human developer can trust.
What Felo Search Unlocks for Antigravity Teams
Installing Felo Search is a small action — copy a folder, commit to Git. The impact touches every agent task that depends on current information.
Competitive Intelligence on Demand
Before Felo Search, an agent asked about competitor pricing would either guess or say "I don't have current information." Now it searches the live web, compares pricing pages, and returns a structured comparison with sources.
Use cases:
- Pricing analysis: "What are the current plans and pricing for [competitor]?"
- Feature comparison: "Does [competitor] support [feature] as of today?"
- Market positioning: "What's the latest messaging from [company]?"
Each answer is cited, current, and ready to feed into the next phase of the agent's task — whether that's a competitive matrix, a positioning memo, or a product recommendation.
Technical Research That's Actually Current
Developers rely on agents for technical research: API documentation, framework updates, security advisories, library compatibility. All of these move fast, and all of them are unreliable if sourced from a static training snapshot.
Felo Search gives agents access to current technical information:
- API changes: "What changed in [API] v3.2?"
- Security advisories: "Are there known vulnerabilities in [library]?"
- Framework updates: "What are the breaking changes in [framework] 5.0?"
- Best practices: "What's the current recommended approach for [pattern]?"
The agent doesn't just answer — it answers with confidence because the information is verified against the current web, not recalled from training data.
Fast Decision Support
The most valuable thing Felo Search does isn't research — it's speed. When a developer or team lead needs a quick answer to inform a decision, Felo Search provides it in seconds, without the developer leaving their IDE or opening a browser.
Quick questions that matter:
- "Has [tool] been updated in the last month?"
- "What's the current status of [service]?"
- "Are there recent reviews of [product]?"
- "What's the latest on [standard/specification]?"
Each answer comes with attribution. Each answer is current. Each answer lets the team move forward without context-switching to a browser search.
The Bigger Picture: Skills as a Capability Layer
Felo Search is one piece of a broader strategy. It's designed to work alongside other Felo skills that fill different gaps in what Antigravity agents can do autonomously.
The full capability stack looks like this:
| Layer | Skill | What It Does |
|---|---|---|
| Live Data | Felo Search | Current web search, cited answers, real-time verification |
| Live Data | Felo Web Fetch | Structured webpage extraction to Markdown, HTML, or text |
| Live Data | Felo X Search | Social signals from X/Twitter — tweets, users, reply threads |
| Knowledge | Felo LiveDoc | Persistent team knowledge base from indexed docs and files |
| Output | Felo Slides | Generate .pptx decks from agent research |
| Output | Felo Content to Slides | Convert notes, docs, and URLs into slide-ready outlines |
| Output | Felo Landing Page | Generate hosted landing pages from text prompts |
Felo Search is the foundation. It's the skill most teams install first, because the knowledge cutoff gap is the most immediate blocker. Once live data works, teams typically add persistent knowledge (Felo LiveDoc) and then output generation (Felo Slides).
Each layer is a folder install. Each layer is auto-triggered. Each layer makes the Agent Manager more capable without changing how developers work.
Getting Started
Installing Felo Search for Google Antigravity takes under a minute:
# Clone the Felo skills repository
git clone https://github.com/Felo-Inc/felo-skills.git
# Copy the skill to your Antigravity skills folder
cp -r felo-skills/felo-search ~/.gemini/antigravity/skills/
# Or for team-wide usage, put it in your project's .agent/skills/
cp -r felo-skills/felo-search /your-project/.agent/skills/
# Commit to Git so the whole team gets it
git add .agent/skills/felo-search
git commit -m "Add Felo Search skill for live agent research"
That's it. No API key configuration inside Antigravity (your Felo API key is configured separately in the skill's environment). No manual trigger setup. The Agent Manager reads the SKILL.md description and activates Felo Search when the task requires it.
The install path matters:
.agent/skills/(inside your project): Use this for team-shared skills. Commit to Git, and every developer's Antigravity instance picks it up on next pull.~/.gemini/antigravity/skills/(global folder): Use this for personal skills that only apply to your own projects.
Most teams should use .agent/skills/ and commit to Git. That way, the skill is part of the project's configuration — versioned, shared, and consistent across the team.
You can browse the full Felo Skills catalog and compare capabilities before installing at felo.ai/skills/antigravity.
Why This Matters
The promise of AI agents isn't that they can plan. It's that they can do — complete real tasks, produce real outputs, make real decisions. Planning without current information is just educated guessing. And in fast-moving domains — software development, competitive analysis, market research — educated guesses are wrong often enough to be dangerous.
Felo Search gives Antigravity agents the one thing no model upgrade can provide: access to what's happening right now. Not what was true at training time. Not what the model thinks might be true. What's actually true, verified against the current web, cited so you can check.
It's a small install — one folder, one commit. But it changes what your agents can do, from "plan with stale data" to "act on current information."
That's the difference between an agent that's a planning toy and an agent that's a production tool.
Explore all Felo Skills for Google Antigravity at felo.ai/skills/antigravity
This post is also available in 简体中文, 日本語, 한국어, 繁體中文, हिन्दी, Français, العربية, Русский, اردو, Bahasa Indonesia, Deutsch, Tiếng Việt, Türkçe, Italiano, ไทย, Español, বাংলা and Português.