Google I/O 2026: What Changes for Brazilian Companies

Google I/O 2026 took place on 19 and 20 May and left a message that is hard to ignore: artificial intelligence has stopped just answering and started acting. There were new models, agents that work autonomously in the background, and tools that generate media from any input. In this summary, I separate what was announced, what is just stage noise, and what really changes for those running an operation in Brazil.

TL;DR — the 30-second summary

• Google I/O 2026 marked the shift from "AI that talks" to "AI that executes" — autonomous agents were the centre of the event.
• Gemini 3.5 Flash became the default model for the Gemini app and Search AI Mode, optimised for long agentic tasks.
• Gemini Omni creates video, audio, image, and text from any combination of inputs, with built-in SynthID watermark.
• Gemini Spark is a personal agent that runs 24/7 in the background and performs actions under your supervision.
• For businesses, the real gain is not in having the smartest model, but in delegating repetitive processes to agents — with governance.

What Google I/O 2026 announced

Google I/O 2026 was, in practice, an entire conference about agents. Sundar Pichai opened the event describing the current moment as the "agent-driven Gemini era", and almost every relevant announcement revolved around software that decides and acts without needing a human clicking at every step.

Some numbers set the tone. AI Mode, Google's conversational search, reached 1 billion monthly users. The SynthID watermark, which identifies AI-generated content, has already accumulated 50 million uses worldwide. And Gemini 3.5 Flash became the default model for both the Gemini app and AI Mode globally — meaning billions of searches started running on it overnight.

Context is worth noting: the agentic shift is not exclusive to Google. May 2026 was a busy month, with Microsoft, OpenAI, and Anthropic also pushing autonomous agents to the centre of their platforms. Google I/O 2026 carries more weight because Google controls Search, Android, and Workspace simultaneously — when it changes the standard, it changes the ground under almost everyone.

The highlights you need to know:

• Gemini 3.5 Flash — new model focused on action and agentic tasks.
• Gemini Omni — multimodal media creation model, starting with video.
• Gemini Spark — personal agent that works 24/7 in the background.
• Google Antigravity 2.0 — platform for orchestrating multiple agents in parallel.
• Agents in Search — agents that monitor the web for you and send actionable summaries.

The full list is on the official Google blog. Below I detail what really matters for professional use.

Gemini 3.5 Flash: the model designed for agents

Gemini 3.5 Flash is the first model in the Gemini 3.5 family and was described by Google as the combination of "frontier intelligence with action". In other words: it's not just a smarter model, it's a model designed to execute long sequences of tasks without getting lost along the way.

What changed compared to the previous generation

Compared to Gemini 3.1 Pro, the 3.5 Flash improves on almost all benchmarks — including coding. The relevant technical point is the focus on "long-horizon agentic tasks": scenarios where multiple agents work simultaneously, perform actions over extended periods, and make decisions continuously. It is exactly the kind of load that broke previous generation models, which lost context or hallucinated when chaining many steps in a row. This focus on action was the guiding thread of the entire Google I/O 2026.

Where you can already use it

Gemini 3.5 Flash is already the default model for the Gemini app and Search AI Mode. For developers, it is available on the Gemini API, Google AI Studio, Android Studio, and the Google Antigravity platform. Google also confirmed Gemini 3.5 Pro — a more robust version, still in internal use, with launch scheduled for the month following the event. For most integrations of a small company, Flash already solves, costs less per call, and responds faster; save Pro for tasks that require heavier reasoning.

Gemini Omni: creating media from any input

If Gemini 3.5 Flash is the brain of agents, Gemini Omni was the creative arm of Google I/O 2026 and deserves its own attention. It is a model that, according to Google, "creates anything from any input" — it combines text, audio, image, and video as input and generates dynamic content as output, starting with video.

In practice, Gemini Omni merges Gemini's intelligence with generative media models. The first released version, Gemini Omni Flash, is already available for subscribers over 18 and can be tested in the Gemini app, Google Flow, and YouTube Shorts. For a lean marketing team, this means producing a promotional video in minutes from a text script and a few product images.

One detail that often goes unnoticed deserves attention: all content generated by Omni carries the SynthID watermark, imperceptible to the naked eye. For companies, this has two sides. The good: traceability and proof of origin. The complicated: if your strategy depends on passing AI content off as "handmade", this path is getting increasingly narrow.

Gemini Spark and the agentic shift

The announcement that best sums up Google I/O 2026 is Gemini Spark. It is a personal AI agent that runs 24/7 in the background — it works even with your laptop closed because it lives in Google's cloud. Built on Gemini 3.5 and the Google Antigravity platform, it performs actions on your behalf, always under your direction.

Gemini Spark entered beta for AI Ultra plan subscribers in the United States the week after the event. The roadmap includes sensitive capabilities: sending messages and emails directly, creating custom sub-agents, and even authorising payments. For corporate clients, there is a version of Gemini Spark aimed at Gemini Enterprise and Google Workspace.

It didn't come alone. Daily Brief organises and prioritises your day with a personalised summary. Information Agents in Search monitor the web 24/7 on specific topics and send already synthesised updates with action capability. And Google Antigravity 2.0 — now with its own desktop app, CLI, and SDK — allows orchestrating multiple agents in parallel, with direct connection to Google Cloud projects for enterprise use.

For technical teams, the most relevant addition is Managed Agents: they provision a remote Linux environment, isolated in a sandbox, accessible via API. In practice, it gives an agent its own computer to work on — without the risk of it messing with your environment. It is the piece that makes long-running agents viable in production, not just in demos.

Comparison of models presented at Google I/O 2026

To avoid confusing the names, here is a summary of the three core models that Google I/O 2026 put on stage:

Model	Main focus	Status at event	Best for
Gemini 3.5 Flash	Action and long agentic tasks	Available, became default	Agents, automation, coding
Gemini 3.5 Pro	Deeper reasoning	Internal use, launch next month	Complex analyses
Gemini Omni Flash	Multimodal media creation	Available for subscribers	Video, audio, and image

The reading is simple: Flash for speed and action, Pro for heavy reasoning, Omni for creation. Most use cases for a small or medium-sized company fit into Flash, which is cheaper and faster — and that was Google I/O 2026's central bet for the market.

What Google I/O 2026 changes for your company

Fancy announcements on stage don't pay the bills. What Google I/O 2026 concretely changes for those running an operation in Brazil is on four fronts.

Customer service and sales

Agents that monitor, respond, and execute change the game of customer service. Instead of a chatbot that only answers FAQs, you now have an agent that checks an order, updates a record, and sends a reply message. This directly aligns with the operating model I advocate in the post about unlimited agents in business WhatsApp: the cost of adding service capacity plummets when repetitive work is taken off the human team's plate.

Education and distance learning

For those working with distance learning, multimodal models change content production. Generating short video lessons, narrating material, and adapting exercises is no longer a studio bottleneck. Combined with good mobile delivery — as I discuss in the post about advantages of a custom Moodle app — it's possible to scale course production without inflating the team. The caution lies in pedagogical review: the model generates quickly, but ensuring the content is correct remains the human expert's job.

Paid traffic and content

With cheap media generation and built-in SynthID watermark, the bar rises. Creative volume ceases to be a differentiator — every competitor will have it. What again separates those who sell from those who don't is offer strategy, targeting, and measurement. Those who only "produced a lot" will feel the squeeze; those who understand that creative has become a commodity and invest in what AI doesn't deliver will come out ahead.

Product development

For software developers, Gemini 3.5 Flash on the Gemini API and Antigravity 2.0 reduce the cost of embedding agents within a product. But this also means your client will start expecting this from you. You don't need to rewrite everything: just identify a flow in your product where an agent saves user time and start there, before the competitor does.

Risks and pitfalls of the new wave of agentic AI

The agentic wave accelerated by Google I/O 2026 brings real gains, but also concrete risks. An agent that acts is useful and dangerous for the same reason: it acts. Before delegating, it's worth knowing the pitfalls.

• Overly broad authorisation — an agent with permission to send emails and authorise payments is an agent that can make costly mistakes. Always start with minimum permissions.
• Chain hallucination — the longer the sequence of steps, the greater the chance of an error in step 3 contaminating step 9. Critical tasks require human review.
• Platform dependency — building the entire operation on a single supplier creates a single point of failure and price. Maintain portability where possible.
• LGPD and data — an agent that reads email and records is processing personal data. In Brazil, the responsibility for this processing is yours, not the supplier's.
• Traceable content — the SynthID watermark makes AI content identifiable. Plan for transparency rather than disguise.

None of this is a reason to ignore the technology. It is a reason to adopt it with process, not in a panic.

How to prepare without falling for the hype

After following technology keynotes for 15 years, I've learned that the costly mistake is rarely adopting late — it's adopting without criteria. A sober roadmap to react to Google I/O 2026 without rushing:

1. Map repetitive processes — list the tasks your team repeats every week. These are the candidates to become agents, not the creative ones.
2. Start small — choose a low-risk process, automate it, and measure the result for 30 days before expanding.
3. Define minimum permissions — every agent starts with the minimum necessary access. Expand only with a track record of correctness.
4. Keep a human in the loop — for any action involving money, contracts, or customers, require human approval before execution.
5. Document and assign ownership — an agent without an owner and documentation becomes technical debt in a few months.

This roadmap applies both to those who will use Google's tools and to those who prefer another provider. The adoption discipline is the same.

Conclusion — the AI that acts is already here

Google I/O 2026 didn't just present faster models. It presented a role change: AI moved from the passenger seat to the driver's seat. For those running a company, the question is no longer "which AI is smarter" but "which processes can I safely delegate to an agent".

The answer doesn't come from a keynote. It comes from looking at your own operation, finding the repetitive work, and testing methodically. If you want help mapping where AI agents make sense in your business — and, equally important, where they don't — talk to the team at Agathas Web. That's exactly the kind of decision we help you make without hype.