I remember first using a macro back in the early 2000s and that sudden magical feeling that a very boring and grunt-like task had suddenly been made dramatically easier. At the time, I was more than a little impressed. Now, for many in our AI/SI industry, the idea of an autonomous execution of tasks is almost mundane.
At Raise Summit, which we attended earlier this year, Kavitha Mariappan, Chief Transformation Officer at Rubrik, told a story of how an agent autonomously rewrote a company’s entire security policy, completely unprompted. There are now even starker examples. A UK government evaluation saw an agent research real people, invent multiple fake personas and use them in an attempted social-engineering campaign to get malicious code merged into an open-source project.
In another evaluation, OpenAI models independently found a zero-day, used it to obtain internet access and ultimately executed code on dozens of Hugging Face servers.
Stories like this will soon be so regular as to become unworthy of commenting on, but that shouldn’t stop us from being in awe of it.
The Quiet Majority
These are some of the scariest examples, and deserve noting, but they are also the exceptions. For every agent that makes the news and causes another moral panic, there are plenty more quietly getting on with their – or your – work. HUMAN Security found that traffic from AI agents and agentic browsers grew by 7,851% in 2025, while human traffic grew by just 3.1%.
And those numbers have kept moving. In October, DataDome reported nearly 25 billion AI agent requests across its network in the third quarter of 2026, up 41.5% from the previous quarter. Traffic from agentic browsers alone jumped 4.7 times between August and September.
They’re good at it, too. AI agents can process far more information than we can and don’t need to sleep, and the best of them are increasingly able to check their own work, catch their mistakes and adjust as they go, with less and less need for a human to step in.
For months now, our team at Theta Network has been fascinated and preoccupied by the changes in the agentic web, and it’s obviously not just us. With the use of AI agents, builders and individuals are able to exponentially increase their output, with multiple agents acting on their behalf and at their direction.
Blockchain Was the Beginning
Our past, present, and future are rooted in blockchain, and won’t be changing any time soon. Not only that, but in our view, blockchain is a key stepping stone to this idea of the autonomous web. One of blockchain’s great innovations was the idea of the trustless web, which removed the need for trusted intermediaries. In practice, this opened the door to people – and eventually software – transacting without having to trust each other or a middleman.
Nick Szabo’s conception of smart contracts then meant that the rules of a deal could be written in code, and only executed when conditions were met. A revolution that enabled all of DeFi, and now millions of agent payments, escrows, spending limits, identity and reputation.
And it's not just the blockchain industry working on this anymore. In September, Cloudflare began testing a way for agents to pay websites and APIs directly through x402. In October, it also showed how agents could use a website's tools directly, rather than having to work their way through pages and buttons. Mastercard has been experimenting with agent payments too, and demonstrated an agent making nine payments across two providers for research and AI-generated images. All under one approved budget, without anyone having to approve each individual payment.
And there's movement on the compute side, too. This month, Google announced Agent Compute Engine, built specifically for running autonomous agents and managing what they're allowed to do. It's another indication of how much infrastructure is having to change to accommodate software that increasingly acts on its own.
Compute Agents Can Find for Themselves
At Theta, we saw this opportunity early. With our GPU network, we understood that compute – which is analogous to the energy source of AI and AI agents – is required for their activity. So allowing our GPU compute to be available agentically would provide a layer of infrastructure that agents could discover and provision on their own. Since August, this is exactly what agents have been able to do on Theta EdgeCloud. Through our API and MCP server, a standard way for AI agents to use online tools, an agent can browse available GPUs along with their specs, regions and pricing, deploy a node, start, stop or restart it, and check its status and logs before reporting back, without anyone sitting at a dashboard. Every API key is tied to a single project, so an agent can only reach the resources it has been given. That matters once it’s software, rather than people, holding the keys.
This is also where our blockchain roots come back in. Part of EdgeCloud’s capacity comes from people around the world renting out their own GPUs through our EdgeCloud Client, and contributors across our edge network are rewarded in TFUEL on the Theta blockchain. Machines owned by strangers, coordinated and paid without anyone needing to trust each other, is the idea blockchain began with. It’s now powering compute that agents can call on directly.
When I recorded that first macro, I decided exactly what happened and in what order. That arrangement has changed. We now set the goal and the limits, and the software increasingly works out the rest, from finding information to paying for what it needs. A lot of that runs on rails the blockchain world spent the last decade building, and on compute that agents can now find and use for themselves. If you’re building with agents, point one at Theta EdgeCloud and see what it can do on its own.
This article was written by Josh Adams, Head of Marketing & PR at Theta Labs.
This blog is part of our thought leadership series, featuring insights from our team. Subscribe to receive more perspectives and updates from us.
