Agents – The Next Generation
For organizations, a situation in which AI evolves from a productivity-enhancing tool into a user of enterprise systems requires serious attention to matters of architecture, security, and governance.
This field is advancing rapidly, and therefore it is worth addressing several important announcements that have taken place since the previous publication.
Blog
Min readIn the previous article, I promised to address the topic of governance in the use of AI based on autonomous agents in the style of Claw. We have seen that major technology companies are embracing this direction and turning it into products that represent the next generation of agent-based AI. These systems will provide solutions that operate independently, communicate with other systems and AI agents, and execute processes autonomously.
At Google I/O, held in May 2026, Google announced its answer to autonomous agent-based AI in the style I described – Gemini Spark. This is an AI agent that operates continuously. It monitors emails, manages tasks, and can independently communicate with third-party applications. At the same time, Google also launched Search Agents, which add autonomous monitoring and scanning capabilities to the search engine for tasks such as price tracking or ongoing information gathering. At the time of writing, Gemini Spark is initially available to Gemini Ultra subscribers in the United States, but it is clear that this technology will become more accessible in the future.
In the Spark demonstration, Google shows how such an autonomous agent can help. Here are a few translated and adapted examples:
User: Track requirements for interior design internships in New Orleans for the summer.
Gemini: Got it. I will continue searching for interior design internship requirements in New Orleans, evaluate your suitability, and provide organized updates.
User: Every Sunday morning, review all emails from the previous week. Briefly remind me of the open topics and suggest an organized task list by priority. Also block time in my calendar for tasks that require deeper work.
Gemini: Recurring task created.
User: Review every email sent to [email protected] – extract the customer name, date, and lead details. Add the request as a new lead and track all open requests.
Gemini: I have successfully created a tracking task for new customer inquiries. In each tracking sheet, I will record the customer name, email address, request date, and a summary of the inquiry…
Google also announced an official response to the governance challenge called Agent Gateway. This is an infrastructure solution that serves as a control point for every interaction involving AI agents within the organization. Unlike the “open door” approach of early projects, the Gateway monitors, logs, and secures every access point, whether it is a request from a user to an agent, an internal request from one agent to another, or a request to a tool with MCP support. The key innovation is identity-based governance. Every agent has a unique and verified digital identity that is managed through a central directory containing only authorized agents and tools. At the same time, the system can delegate authority to security solutions (many of which are sold and implemented by the cybersecurity and distribution companies within the Peax Group), enabling real-time prevention of attacks such as Prompt Injections or data leakage, while also creating semantic governance policies (that is, meaning-based policies) that evaluate the context of an action.
I started with Google because their announcements are very recent, but AWS (whose products and related expertise are offered by the Peax Data division) is also leading interesting solutions in the same area, although, as usual, in a somewhat different way.
Beyond the technologies for agent developers announced over the past year, AWS is also addressing the governance and security challenges involved in operating desktop tools as part of AI agent workflows. This refers to tools that do not have a structured external API or are not available as SaaS products. For Google, accessing emails stored in Gmail is relatively straightforward because they have direct software access. But what happens when part of the workflow requires operating a desktop application (a client application running on a PC)? To address this, AWS announced a very interesting product – Amazon WorkSpaces for AI Agents – currently in Preview. The product provides an MCP interface to a virtual workstation. It allows agents to open applications running on Windows or in a browser, move the mouse, click and drag, type on the keyboard, take screenshots, and more. Other products provide similar capabilities (for example, Manus AI, a Chinese company that was acquired by Meta and whose acquisition the Chinese government is now attempting to block and reverse), but AWS’s solution does not compromise on security and governance. Applications run inside a controlled “sandbox,” full identity and access controls are enforced, all activity is monitored, the infrastructure supports government-grade encryption, and much more.
These technologies are paving the way for the next generation of personal and enterprise applications, which is already just around the corner. The boundary between an enterprise application and an AI agent is becoming increasingly blurred, while interfaces designed for human users are gradually being enhanced by (and eventually may be replaced by) interfaces designed for AI agents.
What about running models locally? Running a language model locally is very easy today. There are numerous tools that enable local model execution, including running a model directly inside a browser or on a mobile phone. Manufacturers of the “processors” used in smartphones and laptops continue to expand the AI capabilities built into their products (I put “processors” in quotation marks because these are complete systems-on-a-chip, not just processors). However—and this is a significant however—efficiently running a model for AI agents, especially those that require deep reasoning and large amounts of information, demands substantial resources of the kind that are expensive to purchase, operate, and cool locally. This is a broader topic, and I plan to address it next time.
Author
Yoel Jacobsen, CTO, Peax
