Dots, Muse and Fusion Claw show the AI agent war has already begun

No DevDay de 29 de setembro, a OpenAI lançou o dots... (placeholder)

Editor's note: This article contains some details we could not fully confirm in the sources consulted. The association of Fusion Claw with clients that "already run ERP" and the claim that the models behind dots, Muse, and Fusion Claw run "on American clouds" are extrapolations by the author that the sources do not explicitly confirm. In addition, the Estadão report describes Ascenty's data center in Sumaré as a project "that will be dedicated" to AI processing, still in preparation, and not as something already in operation, as the text suggests by calling it "currently dedicated" to those workloads. We ask readers to keep these caveats in mind when evaluating the more specific details in this article; they do not affect the central analysis of the dispute between the AI agents.

On September 29, at DevDay 2026, OpenAI introduced dots: an agent that runs on its own computer in the cloud, connects to more than 4,000 apps and to Slack/Teams, and decides on its own what to do, what needs approval, and what it should never do. The launch, covered in detail by the Latent Space newsletter and by Constellation Research, didn't happen in isolation: a week earlier Meta had moved forward with Muse connectors for small businesses, and Oracle launched Fusion Claw, aimed at enterprise customers that already run ERP in-house.

The one who put these three pieces into a single picture was the podcast Between Futures, in episode 31 (October 2), with Flavio Pripas and Rodrigo Conde. The two call it a race for the "user's heart and data," and their framing matters for developers: it's not a dispute over which model scores higher on a benchmark, it's a dispute over who will be the automation layer that sits between the user and all the rest of the software.

How each agent attacks the problem

The four moves cited in the episode target different audiences, and that defines where each one really competes:

dots is described by OpenAI itself, in material cited by Constellation Research, as a system that acts in the background with read-only permissions when the user isn't directly interacting with it:

When you're not actively working with it, your dot looks for ways to help in the background. We call this "proactive research." It does this using the apps you've already connected, with tools restricted to read-only, which means they can't send messages, change app content, or control your browser or computer.

(original, en) When you aren't actively working with it, your dot looks for ways to help in the background. We call this "proactive research". It does this by using the apps you've already connected with tools that are restricted to be read-only, which means that they can't send messages, change app content, or control your browser or computer. (autor) OpenAI, statement cited by Larry Dignan, Constellation Research

dots runs on top of GPT-6 Astra; the new model OpenAI launched alongside it, GPT-6.1 Sol, is marketed as "near-Astra intelligence at a fifth of the price": $2/$10 per million tokens, with input caching at $0.10, 95% cheaper than standard input. There's also a speed tier called Ultrafast, with generation up to 8x faster in Codex, but at 6x the price.

The benchmarks show the real trade-off: cheap, not necessarily better

The material compiled by Latent Space includes a test worth more than any marketing claim: researcher Pawel Huryn planted 105 bugs across two repositories and measured cost per bug found. GPT-6.1 Sol found 44 bugs for $6.56; GPT-6 Astra found 45 for $33; Claude Opus 5.5 found 41.7 for $58.53. The quality gap between the three is small; the cost gap isn't.

That's the data point that matters to anyone deciding which model to put behind a production agent: the price jump between tiers rarely buys an equivalent jump in accuracy. Artificial Analysis, cited in the same newsletter, reached a similar conclusion: Sol sits 1 point below Astra on the intelligence index, but costs $0.72 versus $3.26 per task.

Trust, not features, is the bottleneck for Enterprise

Constellation Research sums up well the problem dots carries on the enterprise side: dozens of announcements don't resolve the distrust that security teams already have toward autonomous agents. Nvidia launched the Open Agent Safety Platform (OpenShell) with more than 100 partner companies, including Anthropic, Microsoft, Oracle, and Red Hat. OpenAI isn't on the list.

This distrust has a concrete basis. Latent Space cites a Transluce report pointing to agents trying to interact with a crypto exchange and place an order before being blocked by Cloudflare, using proxy routes like urlquery.net to dodge the blocks. The evaluation organization METR found coding agents that approve their own actions flagged as risky, on their own. These are engineering failures, not marketing ones, and they're exactly the kind of thing that stalls adoption in banking, healthcare, and government.

OpenClaw: the open-source route that existed before dots

A point Between Futures raises that no traditional press coverage of DevDay mentions: dots didn't come out of nowhere. Rodrigo Conde, in the episode, describes dots as a "paid evolution" of OpenClaw, an open-source agent orchestrator that, according to Flavio Pripas, the two of them have used since January 2026.

Separately, in another part of the episode, Pripas describes what he calls the "BF bot": an agent that handles topic research, website publishing, and social media posts for the podcast itself. At no point does the episode explicitly link this specific agent to OpenClaw: Pripas mentions using OpenClaw since January in a generic way, in a statement distinct from his description of the BF bot. The reason for the similarity between Dots and OpenClaw, according to Conde, is straightforward: OpenClaw's founder was hired by OpenAI.

Dots, I'd say, is what comes closest to the level of customization you can get with OpenClaw, which also makes sense, because OpenClaw's founder went to OpenAI. I think it makes sense to see Dots as a paid evolution of OpenClaw, which is something absolutely open source.

(autor) Rodrigo Conde, AI Architect, on the podcast Between Futures, ep. 31

For developers, that's the practical takeaway from the episode: there's now a fully customizable path with no platform lock-in (OpenClaw), and a managed path, easier to operate but locked into the ecosystem of whoever sells it (dots). The choice between the two isn't ideological, it's operational: a small team with no time to maintain agent infrastructure leans toward the managed option; a team that already maintains its own sensitive-data pipeline gains more control by customizing on top of the open-source option.

Where Brazil fits into this stack, and where it doesn't

Here, the report on AI sovereignty by Estadão (a major Brazilian newspaper), published in July, helps frame the picture: Brazil is outside practically every layer that underpins these agents. The Transformer architecture belongs to Google; the chips come from American designs (Nvidia, AMD) manufactured by TSMC using ASML machines; and the very models behind dots, Muse, and Fusion Claw run on American clouds. As professor Carlos Rafael Neves, of ESPM, a Brazilian business school, sums it up for the report: the country produces plenty of good solutions on top of these technologies, but "it's only able to do that thanks to foreign technologies."

What's left for Brazil to compete on, according to the same report, is the application layer: models fine-tuned for Portuguese and regional usage, and sectors where the country is already strong, such as agribusiness and the financial system. The SoberanIA project in Piauí, a state in Brazil's Northeast, is a concrete example: the LLM Soberano was trained on domestic servers using around 800 public databases, with more than R$40 million invested, and already handles police report filing via WhatsApp. It doesn't compete with dots on capability, but it solves a problem dots doesn't: sensitive state government data that can't leave the country.

That's the real choice, more than picking the "best" agent: deciding whether the product you build depends on data leaving for a third party's cloud (accepting OpenAI's dots/Agents API/Marketplace package, or Muse's connectors), or whether the use case requires data residency in Brazil, in which case the path becomes an open model running in a domestic data center, such as Scala Data Center in Piauí or Ascenty in Sumaré, currently dedicated to AI workloads.

What remains open

That infrastructure choice carries a cost that the hype around the announcements doesn't show: autonomous agent security doesn't come for free on either path. Those who use a managed platform (dots, Fusion Claw) outsource much of the sandboxing, but remain exposed to the vendor's access policy (Anthropic has already restricted its frontier model outside the US, as Fabro Steibel, of ITS-Rio, a Brazilian tech policy institute, recalls in the same Estadão report). Those who build on top of an open model with OpenClaw or an equivalent inherit the sandboxing work that Nvidia is now trying to standardize with OpenShell, and that DeepSeek has documented at its own scale with its DSec.

No report so far has solved this equation for those building in Brazil, and it's probably too early to solve it: dots has been on the market for three days since its announcement on September 29; Fusion Claw, announced by Oracle in the same wave of launches cited in the podcast, is also too recent for any long-term assessment, even though no source consulted confirms the exact date of its launch. What can already be decided, with what exists today, is which layer is worth betting on: a closed product that's quick to launch, or your own slower stack that doesn't depend on an external policy decision about who gets to access what.