Meta, OpenAI, Grok, and Oracle launch 24/7 agents and expose the Brazilian dev's dilemma

The month the agent war got real

In just a few weeks, four companies bet on the same idea: an AI agent that stays on 24 hours a day, with its own computer and browser, acting on the user's behalf instead of just answering questions. Meta launched Muse. OpenAI launched Dots on September 29, 2026, during its Dev Day. xAI got Grok Bot up and running, and Oracle, targeting enterprises, launched Fusion Claw. That's what hosts Flávio Pripas and Rodrigo Conde discuss in episode 31 of the Between Futures podcast.

For developers, this isn't just another consumer product launch. It's a fight over who will control the layer that decides what an AI system can do on its own: read your email, pay a bill, book an appointment, cancel a spam subscription. In short: the product stopped being a chat's answer and became the action executed without you watching.

Four agents, four different technical bets

The four launches don't compete for the same user or solve the same problem. Pripas, who tested Dots closely, sums up the core difference: the easier the setup, the less customization is possible, and vice versa.

None of the four, according to the hosts, currently lets the agent talk to another agent on another platform. It's an interoperability gap that echoes the problem protocols like MCP try to solve between model and tool, except here the bottleneck sits one level up: between agent and agent.

The precedent SAP shouldn't have set

The episode revisits a case discussed in an earlier edition of the podcast: SAP had limited AI agents' access to its corporate APIs, a decision market analysts called a self-inflicted wound. Pripas and Conde note that SAP, Microsoft, and Oracle had been falling behind in this agent race, and that Oracle reportedly pulled ahead precisely by opening its data to Fusion Claw instead of closing it off.

For those building integrations, that's the practical takeaway: the API that treats an agent as a threat loses ground to the API that treats an agent as a customer. Blocking programmatic access may protect data in the short run, but it takes the company out of the game once the usage pattern becomes "ask the agent to handle it."

Build or import: open source's role in this

Since January 2026, the hosts have been running their own open source agent orchestrator called OpenClaw, used to maintain the podcast's website, organize the editorial lineup, and publish social media posts with no manual intervention beyond video editing. It's the concrete proof that it's possible to build an agent layer without depending on Muse, Dots, Grok Bot, or Fusion Claw.

There's a direct connection between that project and Dots: according to Pripas, OpenClaw's founder was hired by OpenAI, and Dots ended up becoming what he calls the "paid evolution" of a project that started out open source. That says something about this moment: the line between building your own orchestration and subscribing to a lab's version is getting more porous, and whoever knows how to touch the code still has a customization edge that none of the four showcase agents deliver.

I think what's happening in the market is fascinating. I think it's a new evolution of AI: it stopped being just a conversational system where you ask questions and it gives you little answers.

Rodrigo Conde, co-host of the Between Futures podcast

What's still left open

The question the hosts leave hanging is practical and gets no closed answer in the episode: once an agent starts executing work on its own, who controls permission, memory, cost, and accountability for what it does? None of the four launches solves this publicly and auditably; each solves it its own way, inside its own closed ecosystem.

According to the episode description, the conversation also touches on the voluntary pact signed between the White House, AI labs, and chipmakers, and on Google's attempt to retake the frontier with Gemini 4 Argon. The available material doesn't detail the terms of that pact or the performance numbers for Gemini 4 Argon, so these two fronts are best treated as context to be explored further, not as settled fact here.

Where Brazil decides to step in

The angle that gives the episode its name is direct: the country can try to compete on models and infrastructure, which requires computing capacity Brazil doesn't have today at a comparable scale, or it can bet on expanding access, adoption, education, and digital government on top of what already exists abroad. For developers, the second option isn't a consolation prize, it's where the real technical work lives.

Building the permission, audit, and cost layer for an agent that runs on top of an imported model is as complex as training the model itself, just cheaper and more accessible for a small team. OpenClaw's example shows that this path is already being walked; the open question is whether the Brazilian ecosystem will treat it as a strategic product or just as a makeshift fix bolted onto someone else's tool.