Quanitzed AI News 26/09: Always-On Agents, Gemini 4 Argon & Pi 1

**TL;DR: This edition of Quantized AI News focuses on three developments—OpenAI’s launch of the always-on agent “Dots,” Google’s strong return to the frontier with Gemini 4 Argon, and the design philosophy behind Pi, a minimalist dark horse among coding agents.** # Quantized AI News 26/09: Always-On Agents, Gemini 4 Argon, and Pi Hello everyone, welcome to Quantized AI News from Soverius AI. I’m Rainer Hahnekamp. This edition covers always-on agents with OpenAI Dots, Gemini 4 Argon, and Pi. ## OpenAI Dots: Always-On Agents Running in the Background OpenAI has launched Dots, a type of agent that runs in the background, integrates with various existing tools such as Slack, and can also carry out tasks autonomously within the permissions we grant it. These features are not free. You need a Pro license, Business Premium, or Enterprise. Unfortunately, Pro is unavailable in most European countries, so your only options there are Business Premium or Enterprise. OpenAI has also given Dots a visual identity to make it more intuitive and tangible for nontechnical users. The core idea is that as long as an agent can access the relevant applications, it can work on all kinds of tasks for us. ### Always-On Agents Were Not Invented by OpenAI OpenAI clearly did not invent this type of always-on agent. The concept entered public awareness at the beginning of this year—it was OpenClaw that made us realize we could have an agent running continuously, without needing a specific user interface. Instead, we could connect to it through existing tools we already use, such as WhatsApp. Later, different variations on this approach emerged, including Hermes and Meta’s Meta Muse. An interesting fact: after OpenAI Dots launched, CopilotKit also released an open-source version of Dots that does bear some resemblance to OpenAI’s version. ### Why Always-On Agents Are Becoming Popular The more important topic is not OpenAI Dots itself, but the bigger picture—always-on agents are becoming increasingly popular. The reasons are easy to understand: - These agents are always online - They integrate into our existing tools, so end users do not need to install anything extra - We do not have to deliberately write prompts, especially when they can discover tasks on their own within the permissions we define If they can complete these tasks satisfactorily, we can certainly call always-on agents a milestone in the widespread adoption of AI. ## Gemini 4 Argon: Google Makes a Strong Return to the Frontier Google was an early participant in AI, contributing a great deal of research and pioneering work. But when ChatGPT appeared, we felt for a while that Google had fallen somewhat behind. They made a comeback with Gemini 3, but then seemed to struggle again to keep pace with the innovation brought by new models from Anthropic, OpenAI, and Chinese companies. Now Google is making another comeback with Gemini 4 Argon. Judging by the launch blog post, the data does indeed show that Google is back at the table—they have returned to the frontier and caught up with the leading frontier labs. ### Release Schedule and Benchmark Performance We currently know little about the hands-on experience because, as of this recording, the model is not yet officially available. Google has followed a process we have already seen from OpenAI and Anthropic: first making the model available to selected partners to ensure its safety. One of those partners is Artificial Analysis, which has already updated its benchmark results. We can see that Gemini 4 is indeed very close to, or arguably already among, the frontier models—it is roughly on par with GPT-6. Its cost per task is also quite competitive. Although the model is not yet publicly available, there is no doubt that Google has made a strong comeback. ### The Advantages of an Established Company Looking closely at Google, we see an established company in a different position from the frontier labs. Google has other business markets and does not face the urgent pressure to lead at all times, so it can take things a little more slowly. But when the need arises, it has enough resources and expertise to build a new frontier model. Another aspect is safety. At other companies in the AI industry—not Google—we often hear and read about internally established rules and procedures. But then problems and incidents occur, and it turns out those guidelines were not actually being followed. Google is in a much better position: as an established company, it has had time for its internal guidelines and culture to mature, allowing it to introduce guidelines that have been validated for safety and are actually followed. Of course, we can never be 100 percent certain, but we do have a degree of trust in Google because it has been around for many years. So overall, in the field of frontier models, having Google in charge is actually quite reassuring. ## Pi: A Minimalist, Pure Dark Horse Among Coding Agents Pi is a coding agent. It is not as popular as Codex or Claude Code, but it is certainly not obscure either. It is a very minimalist, pure coding agent that we consider a dark horse. This is Mario Zechner, the creator of Pi. He gave a talk at AI Engineer Europe about six months ago that provided a very good introduction to what Pi is and how it differs from other coding agents. ### Adapting to Your Workflow Instead of Making You Adapt to It Pi positions itself as an agent that adapts to your workflow instead of making you adapt to it. It consists of four packages: - **AI package**: A cross-provider abstraction layer and context switching between providers - **Agent core**: Just a while loop plus tool calls - **Custom TUI framework**: The author comes from game development, so he built an interface that does not flicker too much - **The coding agent itself** ### A Tiny System Prompt By comparison, Claude Code requires 22,000 tokens just to initialize, while Pi’s system prompt is very small. Later, the industry created a new standard called Skills, which essentially consists of Markdown files. Pi added this feature too, and because it needed to be included in the system prompt, they very reluctantly added a few more lines. The key magic that allows Pi to modify itself is the published documentation—written manually by Mario and an agent—and code examples for extensions. To get the agent to modify itself, you only need to tell it: here is the documentation, and here is some code showing you how to modify yourself by writing extensions. Pi’s focus is extensibility—that is why it can be so minimalist. Whatever feature you want, you can simply tell Pi and it can help you implement it, or you can write a custom extension yourself. ### Four Tools, and That’s It The minimalist philosophy also shows in the number of tools. Pi has only four: **read, write, edit, bash**. That’s it. The smaller the system prompt, the easier things become in terms of startup speed, token budgets, and so on. In fact, these models have already undergone extensive reinforcement training, and they know very well what a coding agent is—because coding agent frameworks are essentially their training environments during post-training. You do not need ten thousand tokens to tell them they are coding agents. They know, because they are coding agents now. What Mario really trusts here is the result of the models’ training. ### YOLO Mode and a Focus on Professional Users When it comes to safety, Pi also defaults to YOLO mode because “my safety needs are different from yours.” Mario does not consider a dialog asking for approval every time Bash is invoked to be a particularly smart safety mechanism. So he gives users enough freedom to build whatever is appropriate for their specific safety needs. This is definitely a coding agent for professionals: letting an agent run freely means taking responsibility yourself. You need to know what safety constraints to apply in a particular situation. ### Version 1.0 and Model Independence Pi has now released version 1.0, with several new features: native MCP support and a built-in safety feature called Codemode. Pi is released by Earendil, an Austrian company, and you can also obtain related services for it. Since Earendil is not a frontier lab, Pi does not have a built-in model like Codex or Claude Code. However, you can use your existing subscriptions with Pi—for example, by logging into Codex or Claude Code and then using their models in Pi. Of course, it also supports other models, such as Kimi and Google’s models, or general connectivity through OpenRouter. What matters is that we have this kind of coding agent that is not tied to a particular model preference. Pi is certainly not the only one, but having these coding agents creates opportunities for open-weight models and lets users try them. These models now run inside coding agents, and those coding agents operate in an environment supported by a community—with plenty of extensions in Pi’s case. Source: Quantized AI News 26/09: Always-On Agents, Gemini 4 Argon & Pi 1 (https://www.youtube.com/watch?v=wTZSNFK2lfs)