OpenAI Dots and Always-On Agents
OpenAI presented its Dots, agents that run in the background, integrate with different tools, and autonomously do their work within the permissions we give them.
Dots are being rolled out to eligible Pro, Business Premium, and Enterprise users. In Europe, Pro access is not available yet.
OpenAI also gave Dots a visual character, which makes the assistant more tangible for non-tech users.
Dots are just there. They do their work and communicate with us through familiar channels like Slack and Teams. So end users don't have to open a specific agentic application for every task. These agents are available where we already work. This makes them much easier to work with, especially if they proactively find useful tasks and complete them to our satisfaction.
That doesn't mean unlimited autonomy. Connected background research is read-only, and further actions depend on permissions and approvals.
The idea of having always-on agents is not new, and OpenAI is definitely not the inventor. Always-on agents became more widely known with OpenClaw at the beginning of this year. Messengers were a primary interface, and background work and scheduling were part of the idea.
We have also seen projects like Hermes and Meta Muse. OpenAI's Dot is another example of this approach. CopilotKit released OpenDots, an open-source alternative, shortly after Dots was announced.
Analysis
For us, the bigger topic is not just OpenAI Dots. It is that always-on agents are becoming a new technology layer. They are there, they work in the background, and they reach us through channels we already use. We don't have to explicitly prompt them for every single step or install a separate application. If those agents prove they can find and finish their work to users' satisfaction, this could be another milestone in bringing AI to a wider audience.
Gemini 4 Argon
Google was an early player in AI and has contributed a lot of research and pioneering work. But when ChatGPT came out, it seemed that Google was losing ground and couldn't keep up. They had a comeback with Gemini 3, but to us it still looked like they were struggling to keep up with the rapid progress we saw from Anthropic, OpenAI, xAI and the Chinese vendors.
Now, so it seems, they have a big comeback with Gemini 4 Argon. Google has just made Argon available to Google AI Pro and Ultra subscribers. Since access has only just opened, we don't have hands-on experience with it yet.
Google has already released a blog post, and evaluators such as Artificial Analysis have had access.
Google also published a benchmark comparison in its announcement, which we can look at here. These are Google's reported results, with the setup explained in its evaluation methodology.
Analysis
Looking at the benchmarks Google published, Argon looks like a big comeback. Of course, these are Google's reported results, but they show why this release is worth watching.
We see that Google, as a mature company, doesn't have to rush or throw everything in to be at the frontier. They have other markets as well. Still, they undoubtedly have the resources to develop frontier models.
The other aspect that speaks for Google is security. Looking at the incidents at the labs and what has become public over the past few months, we see a gap between what these labs say about security and what happens in practice. This is our interpretation of those incidents, not a measured comparison of the models' safety.
On the other side, we have Google, with established guidelines, rules, and quite a lot of company experience and knowledge about security. From our experience with Google, security plays a different, much more central, role there.
And that's why we can expect, although we don't know for sure, that when it comes to security concerns, models may be in better hands at Google than at some of the other labs. That is our analysis, not a guarantee.
Pi 1
Pi is a coding agent, less popular than Codex or Claude Code, but far from unknown. It is seen as a more purist, minimalist agent. One could call it the underdog of the coding agents.
In its earlier releases, it had only read, write, edit, and execute tool calls: no built-in MCP, no special web search, no built-in subagents, or a hardened permission system. But that was on purpose.
Pi is extremely extensible, so professional developers can customize and extend it exactly how they need.
Pi has been released in version 1 under Earendil. Its primary surface is the CLI, but community desktop interfaces are available as well.
It is not bound to a single model vendor. It supports different providers, using subscriptions where the provider permits that integration, or API access and separate billing otherwise. We should not confuse a working login integration with permission to reuse every provider's subscription.
Pi 1 now also supports MCP and Codemode. Codemode lets the agent use small scripts to combine tool calls and process their results. That sandbox is not the same as sandboxing everything the coding agent can do on your machine.
Alongside Pi 1, Earendil released Pi Durable, an experimental framework for building long-running agents that can resume their work after interruptions. It is a foundation for always-on agents, not a ready-made personal assistant.
One last thing we would like to add from a personal perspective. Like Soverius AI, Pi is also Austrian. It was created by Mario Zechner, who joined Earendil, the company co-founded by Armin Ronacher and Colin Hanna.
Analysis
It is important that we have coding harnesses which are not bound to a very specific vendor. This lets us use different models with mature coding agents, plus a community, plugins, extensions, and more.
We also think it is good when the labs support users reusing their subscriptions in other coding agents. OpenAI provides an official route for participating clients. That is the direction we would like to see more vendors take; not every vendor currently supports it on the same terms.
Read the original article on Soverius AI.
Top comments (0)