Brain and Body: A Single-Screen Agent Setup with GPT-6 Astra on Hermes

The typical day of an AI user starts on screen: a chatbot in one tab, a coding agent in another app, a lighter model in a third window, a Hermes desktop in some corner. Constant clicking, lost focus, unfinished tasks. The host sums up his answer in one line: Astra is the brain , Hermes is the body. One thinks deeply, the other goes and finishes the work, and it keeps working while the user does something else. The narrator introduces himself as the digital twin of an agency chief executive, says the executive handles client work while he delivers the updates, and invites viewers to comment.

First the brain: GPT-6 Astra was announced on 3 September 2026 as, according to OpenAI, the most capable model the company had broadly released to that date. OpenAI positions the model as state of the art in computer use , browsing, software engineering, cybersecurity, science and professional work. The technical identity is crisp: roughly a 1.05-million-token context window, 128 thousand tokens of maximum output, text and image input with text output, a knowledge cutoff at the end of April 2026, and adjustable reasoning effort from low to max. Access opened in stages, reaching developers through the API, Azure and Bedrock after the Plus, Pro, Business and Enterprise tiers.

According to OpenAI the shared measurements are bold, so they should be read with care: 98 percent on the FrontierMath Tier 4 test, 99.9 percent on the ARC-AGI 3 measure and 100 percent on the ExploitBench cybersecurity test. On computer use, Astra completes 72.6 percent of OSWorld 2.0 simulation tasks in about forty minutes while the previous flagship stays at 65.7 percent in about seventy-five minutes, which the company presents as nearly half the time per task. The everyday equivalents are more concrete: filling in online forms, updating customer records, organising a calendar, researching and drafting summaries in email or a document editor, inspecting data and producing plots, building a site and running front-end quality checks. The in-house example the host relays sticks in the mind: the developer and marketing teams built a launch video from three hours of multi-camera raw footage together with Codex, and it reached 550 thousand views in four days.

Then the body: Hermes is presented by the host as born from a research community and billed as the world's most used open-source agent harness. Open source means free and modifiable. According to NousResearch docs the core holds a closed learning loop: the agent keeps preferences and project context in persistent memory files, searches past conversations in full text, writes reusable skill documents from hard jobs it solves, and improves them in use. Scheduled automations, isolated subagents, tool connection over MCP and named helpers working together in group chats are described in the same docs. According to hermes-agent.org the package is completed by automatic skill creation, a gateway joining Telegram, Discord, Slack, WhatsApp, Signal and the terminal, unattended morning reports and backups, parallel work streams and full browser control. In short, the body is an operating layer that runs on schedule even while its owner sleeps.

Connecting the two is described as easier than expected. Hermes allows a code login with a ChatGPT account, and anyone already using Codex on their machine can inherit the same login, so no interface key is needed to start. Who holds Astra is a tier list: Plus, Pro, Business and Enterprise. According to Aiprofitboardroom the timing of the pairing matters: the interface rollout lands days after the chat tiers, so the agent side is the last to meet the brain. The same article names three signals to watch: the model id appearing on the developer dashboard, its listing on the OpenRouter side, and opening a dedicated Astra profile. According to Myclaw the rollout moved in stages, limited organisations in the safety programme first, then paid tiers and the interface. A cloud subscription needs interface use to run inside, the current chat plan is not enough. The host's demo is one sentence: Codex was opened, Astra was picked as the model, a new profile inside Hermes was requested, and the setup happened.

Why is changing only the model not enough? Because inside Hermes each profile means a separate agent, with its own settings, its own keys, its own personality file called soul.md and its own memory. In the host's setup one profile holds Astra, another a different model and separate profiles hold local models; as each new model arrives a new profile is added and switching brains takes seconds. There is more: each profile can now present itself as a distinct assistant with its own display name, assigned part, model, stored context and avatar, routing tasks through a shared agent inbox . This discipline is the practical way to run experiments without contamination, racing the new brain against the old one on the same job while the current setup stays untouched.

One screen: the right brain for the job

The knot of the video is why Astra should move inside Hermes when Codex alone is strong. The answer is blunt: the Codex screen barely changes, no custom button or custom tool can be installed inside it, and it does not adapt to a personal way of working. Hermes is open source, so it bends to taste, and the host's agent operating system is the product of that bending: mission control running on his own machine, opened in the browser like a site. The old days were crowded, the new day is one screen: Codex a click away when wanted, Hermes a click away when wanted, and a brain change in one move. Astra for heavy and tangled work, no need for the strongest brain on light work. OpenAI's Luna model passed similar training to Astra but was built for speed, and even Microsoft suggests leaving summaries and triage to Luna and deep thinking to the big model. The examples are crisp: five social posts for coaching calls go to the light model, a monthly content plan goes to Astra. The right brain is picked and the screen is never left.

The first layer built on the panel is the voice agent, Hermes Apollo. The analogy is Jarvis from the cinema; speech goes out loud, the answer comes back by voice, and it feels like ordinary chat because ChatGPT real-time voice is used. Everything spoken and everything produced during production appears on screen. The host gives the example while walking: asking for a welcome message that tells new community members where to start is enough, no typing needed. Hands free, work moving.

Scheduled panels and one-click production

The next layer is the daily-used Oracle. A custom screen pulls trending topics in AI, runs on a 24-hour schedule and stays current. Hermes was built for exactly this kind of work that runs on schedule even during sleep. Scheduling exists on the Codex side too, but side by side the Oracle side wins on looks and function. The button makes the difference: on spotting a hot topic the WordPress Publish button is pressed and one click produces four SEO articles for four separate sites. When a new AI tool lands, four pieces on how members save weekly hours are ready in one move. That button does not exist in the Codex scheduler, because the team built it themselves. The message is plain: the difference from an ordinary downloaded app is building your own software tool that runs on Hermes with Astra as the brain.

The last layer eases the idea side: competitor watching, what trends in the field, and content suggestions grounded in the current moment. Instead of staring at a blank page, ready ideas open up, for instance a video idea on five AI workflows members set up in their first week. The design is clean and fun to use, and that detail matters: a loved tool gets reopened. On the visual side the Astra difference sharpens: another subscription cannot be added inside Hermes, yet the chat robot can generate images, so doors open. First Image Studio: a small gallery where a typed request turns into images stored in one place. Then the favourite, Ad Studio: an ad topic is typed and the full ad arrives with image, copy and creative direction; one click regenerates it when disliked and every ad rests in a library. The example request is concrete: an ad for the weekly coaching call showing a business owner setting up an agent, image and copy ready in seconds. The weekly routine shrinks to buttons: SEO content here, ad creative there, video ideas in the other corner. Which app or agent is gone as a question; one question stays: which job saves time. The founding story of Ad Studio is three messages: an ad-producing section was requested from the chat robot, three rounds were exchanged, and a working studio was born. The whole setup, meaning the ready zip file, a one-hour course, four coaching calls a week and the member map, folds into one sentence: all of it is offered through membership via aprofitboardroom.com.