Understanding SI, Part 3: SI Agents Ranked, With the Computers They Keep

Every large super intelligence (SI) lab now sells the same thing, an agent with a computer of its own that keeps working after you close the app. Anthropic opened a research preview of Claude Duet on October 21, 2026. It joined Grok Bot, which SpaceXSI launched in beta on August 11, Meta’s Muse, which Axios reported on its September 8 launch, Google’s Gemini Spark from May 19, and OpenAI’s Dots from September 29. Microsoft has its own in private preview, and smaller companies sell the same idea through text messages or on your own desk.

This site now writes SI where it used to write AI, after the executive order the Grok section describes. Grok Bot is the reference every other agent is measured against here. An agent that never closes needs a machine, a memory, and a long context, and all three end up on DRAM, HBM, and flash.

What an agent is in 2026

A chatbot answers and stops. An agent takes a goal, breaks it into steps, uses tools to carry them out, and keeps going until the job is done or it needs a decision. Underneath sits the same kind of model that powers a chat. A harness wraps the model. That program hands the model its tools, runs the tool calls it asks for, feeds the results back, and decides when to stop. How the model itself works is covered on its own page. This one is about what has been built around it.

Figure 1 reads from left to right. Your request goes to the harness, the harness sends the model’s tool calls out to apps and the browser, and anything that cannot be undone stops at the gate. Underneath all of it is a computer that belongs to the agent rather than to you.

Five parts every agent here shares

SpaceXSI’s designers reduced their own agent to five objects, and the list fits the whole field. Grok Bot’s design post names persistent agents with their own memory and tools, chats to talk to them in, prompts that can be saved as skills or scheduled as routines, tools that reach other software, and the artifacts the agents produce.

Under those names, every agent in this guide has the same machinery:

• A model that plans and writes the tool calls.
• Tools: connectors to apps such as Gmail or Slack, a browser, a terminal, and sometimes a phone line.
• A computer to run them on. Since mid-2026 that is a virtual machine in the maker’s cloud, so the work continues when your laptop is shut.
• Memory that outlasts one conversation, plus a way to search past work.
• A gate between the agent and anything it cannot take back, such as sending, paying, signing, or deleting.

Why the computer moved to the cloud

This year’s change is the cloud machine. Claude Duet has a virtual machine with its own browser, so it does not need your phone or laptop to be on. Muse keeps both the agent and the person’s data in a dedicated virtual machine Meta calls Muse Secure VM. Each Grok Bot computer has a desktop, a filesystem, a terminal, and apps, and you can take it over like a remote desktop, as SpaceXSI’s getting-started guide explains. Gemini Spark is a cloud-based agent that keeps working when the laptop is closed or the phone locked. Each of OpenAI’s Dots has its own cloud computer and browser, and Tab has its own computer and browser behind its text messages.

A machine that runs all day changes how the agent gets its work. A chat starts when you type. Grok Bot’s routines start on a schedule or an event, and Spark takes recurring tasks and triggers. Claude Duet checks your connected accounts every few hours, because checking nonstop would use up the plan’s limits much faster. Every look costs model time, so how often an agent wakes is a trade against what it spends.

The ranking

The tables hold every agent, as of October 2026, that a person can talk to in plain language, that acts in other apps on its own, and that keeps working with the app closed. That rule takes in six agents from large companies and four from younger ones. Perplexity’s cloud agent, Computer, shares a row with Personal Computer, which brings the same agent to your own machine. Coding agents and local harnesses get their own section, because they do a different job.

In Figure 2, Grok Bot, Dots, and Tab carry the most published abilities, and Tab is the only agent whose own maker promises a phone line; Instinct’s comes from press reports. A small dashed ring is not a no. It means the maker hasn’t said, which is common for the startups.

Each fact in the tables is a maker claim, what the company says on its own pages, except Instinct’s row and Muse’s prices, which come from press reports.

Why there are no scores

A benchmark measures a model, and these are products. Claude Duet runs on Claude Opus 5.5 and Dots on GPT-6 Astra, while SpaceXSI says only that it trained Grok 4.7 to understand the Grok Bot harness. A model benchmark says nothing about connectors, approval flows, or cloud machines. Tab and Instinct don’t say which model they use. So the tables give facts per agent, the ranking is an order with my reasons, and a model benchmark appears only where the model maker or an independent lab publishes it. I didn’t want a score that looks precise and isn’t.

The tables

The first table covers what each agent runs on and how it asks before acting. The second covers who can get it and what it costs. Prices and availability change month to month, so both tables are dated October 2026.

Source: each maker’s launch post and help pages, linked in the first column; Instinct from press reports. Tab, Poke, and Instinct do not publish the model they use.

Source: grokbot.com, Axios, Anthropic’s Duet help page, OpenAI’s Dots help page, poke.com, and the launch posts above. A blank cell in either table means the maker is silent on it, as with Grok Bot’s model and Poke’s and Instinct’s approval steps.

My order, and why

Three things set the order. One is how far the agent reaches into other software, another is whether it keeps working with nobody watching, and the third is how much control you keep.

Source: each agent’s own launch post and help pages, linked in its section below. The order is my judgment from those facts.

The top two are close, and the order among the rest would change with one good month for any of them.

Grok Bot against the field

Grok Bot is the agent to measure the others against, because it is the most complete version of the idea and because its maker publishes the most about how it works. Grok’s maker has had three names this year. SpaceX acquired xAI on February 2, 2026, and the business took the name SpaceXAI on July 6. On September 29, Executive Order 14434 told federal agencies to say “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI.” Elon Musk then said SpaceXAI would become SpaceXSI, as The Next Web reported, and the company’s own posts now carry the SpaceXSI name.

The product

SpaceXSI launched Grok Bot on August 11 as a team of always-on agents that have their own computer and keep working around the clock. It launched for SuperGrok and Cursor subscribers on desktop and iOS, with its own usage allowance separate from those plans. Three things set it apart from the others:

• Each Bot works in a cloud desktop with a filesystem and a terminal, so it can use websites that have no clean connector, and you can open the screen and take over when it gets stuck on a login, and a Bot can also run jobs on your own connected Mac, such as a local server that keeps running after the command returns.
• Bots message each other and coordinate in group chats. The design post sets practical limits of roughly 50 Bots per account and six per group chat.
• Tools and skills live at the account level, while memory and routines belong to each Bot, so a finance Bot does not carry the legal Bot’s history.

The same design has a cost, and the getting-started guide spells it out. If you log in to a site with one Bot, every other Bot can reach it too, because they share the computer. Plain Grok has a lighter version of the schedule. Automations on grok.com run a job on a schedule or when an email arrives, without a computer of their own. A separate review agent checks each proposed Grok Bot action and can allow, block, or escalate it, steered by rules you write in plain language and by allow and block lists. It’s a model checking a model, and the guide admits as much. You are trusting it to follow what you wrote.

The model under it

SpaceXSI’s current model is Grok 4.7, introduced on September 21, 2026, at $2 per million input tokens and $6 per million output tokens. By SpaceXSI’s account, Grok 4.7 was trained to understand the Grok Bot harness natively, though no Grok Bot page names the model a Bot runs on. SpaceXSI’s API documentation lists a 500k-token context window and a May 2026 knowledge cutoff.

SpaceXSI’s launch post compares Grok 4.7 with GPT-5.6 Sol and Fable 5.1, on benchmarks SpaceXSI chose. OpenAI had already released GPT-6 Astra on September 3, the date of its safety overview, 18 days before Grok 4.7, and the table leaves it out.

Source: SpaceXSI, “Introducing Grok 4.7,” benchmark table. All rows are maker claims.

An independent index brings in the newer models. Artificial Analysis publishes one Intelligence Index score for each model on its leaderboard. Figure 3 takes each agent maker’s highest-scored model from that leaderboard on October 22, 2026, plus the highest-scored open-weight model.

Grok 4.7 scores 46, 12 points below Claude Opus 5.5 at 58. GPT-6 Astra, Gemini 4 Argon, and Fable 5.1 tie at 53, and Meta’s Muse Wonder Full scores 56. The top model is not always the one in the agent. Meta introduced Muse Wonder on October 12 and says it will soon power more of Muse, which launched on Muse Spark. The best open-weight model, Xiaomi’s MiMo-V2.6-Pro, ties Grok 4.7 at 46.

Where Grok leads and where it trails

In SpaceXSI’s own table, Grok 4.7 ties GPT-5.6 Sol on long terminal work and trails Fable 5.1 by 20 points there, trails Fable 5.1 on long coding tasks, and roughly matches it on office work. It leads clearly on one legal benchmark. The strongest column is price. Fable 5.1 costs five times as much per input token and more than eight times as much per output token. An agent running routines all day pays $6 per million output tokens on Grok 4.7 against $50 on Fable 5.1, and over a month that gap outweighs a few benchmark points. On the independent index, though, Grok 4.7 trails the top models of Anthropic, Meta, OpenAI, and Google.

My read is that Grok Bot leads on the product, while its model trails. Its shared computer, many-agent design, and routines are further along than anything else here, and SpaceXSI’s model is good and cheap rather than the best. The open-source harness helps too. SpaceXSI open-sourced its Grok Build coding agent on July 15, 2026, and says it can run fully local against your own model.

The other agents in the table

Beyond Grok Bot, the field splits by where the agent’s computer lives. Most run on a virtual machine in the maker’s cloud, Perplexity offers one agent in the cloud and one on your own desk, and the open-source harnesses run on your own GPU.

Figure 4’s left panel is the crowded one, and Perplexity appears in two panels. Perplexity Computer, introduced on February 25, 2026, is web-based and runs tasks in the cloud. The other two panels keep the work on hardware you own. A local harness runs the model on your own GPU, and Perplexity’s Personal Computer can run a whole task on the device on Windows and Linux, while on a Mac it splits each task between cloud and local models.

Dots

OpenAI introduced Dots on September 29, 2026, as always-on agents in ChatGPT, powered by GPT-6 Astra. Each Dot gets its own cloud computer and browser, which you can open at any time to inspect its work, and you can give it permission to use your laptop too. It reaches over 4,000 apps through plugins. You can message it in ChatGPT, Slack, or Teams and call it in ChatGPT, and texting is a limited beta for Pro users in the US.

Its controls are close to Grok Bot’s. When you aren’t working with it, a Dot does what OpenAI calls proactive research, using read-only tools that cannot send messages, change app content, or drive the browser. Auto-review checks any action that could touch your accounts against your own rules, and some tasks, such as changing a password, always stay with you. OpenAI’s help page puts the first Dot in the Pro and Business Premium plans at no extra cost, with Pro excluded in the European Economic Area, Switzerland, and the UK at launch, and offers Enterprise workspaces a beta their admins switch on.

Claude Duet

Claude Duet is Anthropic’s agent app, built around Claude Opus 5.5 and aimed at professionals and small-business owners. It sends a morning brief, keeps watch over connected apps, and takes larger jobs such as dashboards or forecasts. It can use over 6,000 connectors and plugins in the Claude Marketplace, and pays through Link by Stripe with a virtual card.

Its limits are written down. The help page says you cannot choose the model, cannot watch or take over its browser, and that it does not make phone calls or text other people. Some actions, such as trading securities, are off limits even with your approval. Its background work counts against the plan’s five-hour and weekly limits at the same rate as chat, which is why it checks in every few hours rather than all the time.

Muse

Meta built Muse for people with no technical background, reached in its own app or in WhatsApp, and powered by Meta’s Muse Spark model. Its security design is the most specific in the field. The agent and the person’s data live in a dedicated VM, and a separate Sentinel agent on the same machine must approve anything Muse sends to the internet. Meta also plans a Confidential VM later in 2026, encrypted with a key only the user holds. At Connect, Meta said Muse is coming to its smart glasses in the coming months and will get its own email address.

Gemini Spark

Gemini Spark runs on Gemini 3.5 and uses Google’s Antigravity harness, and it works closely with Gmail, Docs, and Slides. Google’s examples are recurring jobs, such as parsing each month’s card statement for new fees or turning a school’s emails into a daily digest. It asks first before spending money or sending email. At Google I/O, the company’s developer conference, Google said Spark would go first to trusted testers and then, as a beta, to US Google AI Ultra subscribers.

Microsoft Autopilot

Microsoft introduced Scout in June as its first always-on agent for Microsoft 365, built on the open-source OpenClaw, and renamed it Autopilot on September 25. Each Autopilot has its own identity, memory, computer, and workspace inside a company’s tenant, works through Teams and Outlook, and keeps going while you sleep. Sensitive actions can require a person to sign off, and company data-protection rules apply before anything is sent. It is in private preview, so it is a work tool most people can’t try yet.

Poke

Poke lives in your texts like Tab, through Messages, WhatsApp, or Telegram, messages you first instead of waiting, and runs automations in the background. Poke’s maker joined Cognition on July 23, 2026. Cognition makes the Devin coding agent, calls both products a bet on always-on cloud agents, and counts more than 100 million messages with Poke in the three months before. It has a free plan, with paid plans at $19 and $199 a month.

Perplexity Personal Computer

Perplexity began rolling out Personal Computer on April 16, 2026. It is the odd one out. It runs on your own Mac or PC, works across your local files and apps as well as the web, and stays on only while that machine does. Perplexity’s help page suggests a Mac mini for always-on use, and on supported Mac, AMD, and NVIDIA hardware it can run some tasks with a local model.

Tab

Tab lives in Messages and WhatsApp. Behind the thread it has its own computer and browser, and it is the one agent here whose maker says it will phone a business, wait on hold, and report back. Nothing is paid until you approve it. TechCrunch reported that Tab came out of stealth at a $300 million valuation, and the company is still a private beta.

Instinct

Instinct sells a personal agent that texts and calls businesses for you. It raised $1 billion in Series C funding at a $10 billion valuation from Sequoia, Benchmark, and Coatue, by Pulse 2.0’s report, which also says it is still in early access. Its own site was not reachable for this guide, so it stays at the bottom of the order until there is more to judge.

Local agents: models, runtimes, and harnesses

Everything above runs in someone else’s data center. The same parts can run on your own PC, given an open-weight model, a runtime to serve it, and a harness to make it an agent.

Harnesses

OpenClaw is the open-source, local-first agent the others keep pointing to. It installs on macOS, Linux, WSL2, and Windows, runs a local gateway, and links chat channels such as WhatsApp, Telegram, Signal, and Discord. SpaceXSI lets SuperGrok and X Premium subscribers use their plan inside OpenClaw. The open-source lab Nous Research makes Hermes Agent, which works with any model provider or local endpoint, keeps its own memory and session search, schedules jobs, and runs on seven terminal backends from local to Docker and SSH.

For coding, Grok Build is open source and can point at local inference, and Claude Code’s CLI and IDE versions also support third-party model providers. Several more open-source harnesses will drive a model on your own PC:

Models and runtimes

Ollama and llama.cpp are the usual runtimes on a PC, and LM Studio is an app that downloads local models and runs them for chat or agent work. The model has to fit, and an agent needs a long context. Ollama recommends at least 64k tokens of context when OpenClaw runs on a local model. My reading is that the agent has to hold the task, the tool results, and the conversation at once.

OpenAI’s open-weight models show the sizes involved. gpt-oss-120b runs within 80GB of memory and gpt-oss-20b needs 16GB. The 120b model has 117 billion parameters but activates 5.1 billion per token, and the 20b model has 21 billion and activates 3.6 billion, which is why they run faster than their size suggests.

Open models built for agents

Ollama offers three open models built for agents as downloads for a single PC. NVIDIA’s Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts model that activates 3 billion per token, made for always-on agents in harnesses like OpenClaw and Hermes Agent. Meta’s Muse Glimmer is a 30-billion-parameter model under the Apache 2.0 license, tuned for tool use, long tasks, and recovering from failures. Poolside’s Laguna XS 2.1 has 33 billion parameters with 3 billion active and is aimed at agentic coding and long jobs on a local machine. Ollama’s page for each gives a one-line command that starts it inside Claude Code, OpenCode, or Hermes Agent. Reflection’s Beam is also an open model aimed at agentic tool use, but at 501 billion parameters Ollama serves it only from its own cloud.

The independent index shows the gap to the big models. Of the models I ran, Qwen3.8 27B scores highest at 34, and gpt-oss-20b scores 9, against 46 for the best open-weight model, MiMo-V2.6-Pro. The leaderboard I read listed no score for the three agent models.

What a long context costs on one GPU

On October 22, 2026, I loaded eleven models in Ollama 0.33.3 on my RTX 4090 (24,564MiB) at four context lengths, one load at each. They were the three agent models above and eight other open models, each the default download from Ollama’s library. Ollama’s status report shows how much of each loaded model sits in GPU memory and how much it put in system RAM. At 32,768 tokens I also timed three answers of up to 300 tokens to the same prompt and kept the middle one.

Five of the eleven spilled into system RAM at 131,072 tokens. Qwen3-Coder 30B ended the largest, at 31.9GB from 19.3GB at 8,192 tokens, with 8.9GB of it in RAM, and DeepSeek R1 8B, a 5.2GB download, grew the most, to 25.7GB. In Ollama’s report, gpt-oss-20b stayed at 12.9 to 13.0GB and Gemma 4 26B at 17.4 to 17.5GB at every length. That report is Ollama’s own estimate, though. Even so, the card’s total memory in use, as nvidia-smi read it, rose by 2.7GiB for gpt-oss-20b and 2.4GiB for Gemma 4 26B from 8,192 to 131,072 tokens. NVIDIA’s new model spilled at every length. Its default download is 25GB and it loaded at 25.6GB, too much to fit beside the memory already in use on the card, so 4.3GB of it sat in system RAM even at 8,192 tokens. The 19GB size Ollama also lists for it is an MLX build, which I didn’t test.

The numbers are my own measurement in Ollama 0.33.3 on an RTX 4090 on October 22, 2026, and each speed is the middle of three answers.

Speed didn’t follow size. Laguna XS 2.1 and Qwen3-Coder 30B both topped 170 tokens per second, while Muse Glimmer, which stayed entirely on the GPU, managed 36, the slowest of the eleven. Nemotron 3.5 Lightning still reached 96 with 4.3GB of it in system RAM. A longer context costs GPU room first, and the GPU’s memory decides where the spill starts. An earlier pass the same night, with one answer per model, came out between 6% slower and 64% faster depending on the model, so these speeds give the order of things and are too few for a benchmark.

Safety: approvals, review agents, and prompt injection

An agent reads email and web pages written by strangers, and those pages can carry instructions aimed at it. Anthropic’s Duet help page calls this prompt injection and says no defense is perfect, which is one reason consequential actions need your approval. Every agent here whose maker describes its safeguards answers it the same two ways, with a second check before anything leaves and a human approval for what cannot be taken back.

The second check takes different forms. Grok Bot uses a separate review agent that allows, blocks, or escalates. Muse runs a Sentinel agent that must approve anything sent to the internet. Dots run auto-review against your own rules and keep their idle research read-only, and Autopilot can require a person to sign off on sensitive actions. Claude Duet launched with a classifier designed to reject requests to share the strictest tier of information, including credit card numbers and government ID. Payments in Duet and Muse go through a one-time card, so the agent never holds the real card number.

The weak spots are just as specific. Grok Bot’s shared computer means one login opens a site to every Bot on the account. Duet won’t trade securities or move money between accounts even if you approve it.

What agents ask of memory and storage

A chat can forget you when the tab closes. An agent can’t, and everything it keeps takes a share of DRAM, HBM, or flash.

A computer for every user

Claude Duet, Muse, Grok Bot, Dots, Autopilot, and Tab each give the user a computer in the cloud, and Gemini Spark runs in Google’s cloud. None of the makers publishes how much memory or disk that machine gets, so on October 22 I asked the agents to run nproc, free -h, df -h, and nvidia-smi on their own machines and report what came back. Grok Bot reported 8 Intel Xeon vCPUs, 15GiB of memory with 11GiB in use, and a 128GiB disk with 14GiB used. Muse reported 2 AMD EPYC vCPUs, 7.7GiB of memory with 7.2GiB in use, and a 7.5GiB root filesystem, and listed no other disk. Tab declined, saying “the specs of my machine are my maker’s kitchen.” Neither machine showed a GPU. These are the agents’ own reports from one night, and Grok Bot added that its computer had been rebuilt the night before, so the numbers can change.

Each one still has to hold an operating system and a browser profile, plus the files the agent makes, the credentials it stores, and memory that lasts for months. Muse keeps the person’s data in the VM itself. Grok Bot keeps memory per Bot, up to roughly 50 Bots per account. While a machine runs, that state sits in DRAM, and while it waits, it sits on flash.

How often it wakes up

Duet checks in every few hours because each check costs model time against a usage limit. That one design choice sets how much compute and memory an always-on agent uses when nothing is happening. An agent that checks every few minutes holds its context warm far more of the day than one that checks every few hours.

A long context is a big cache

A transformer model keeps a key-value (KV) cache that holds the attention state for each token it has read, and the cache grows with the context. Why each conversation needs its own cache is in the explainer on how models know things. Agent contexts are long. Grok 4.7 takes 500k tokens, and Grok 4.6 takes 524,288 on Google Cloud. The open Qwen3-32B model shows the scale. Its configuration lists 64 layers, 8 key-value heads, and a head dimension of 128, so at 16-bit precision the cache takes 2 × 64 × 8 × 128 × 2 bytes = 262,144 bytes, or 256KiB per token. It comes to 10GiB at the model’s 40,960-token maximum and 16GiB at 65,536 tokens, before the weights. It’s my estimate, worked from the published configuration.

Figure 7 stops at 65,536 tokens, the context OpenClaw asks for and beyond this model’s 40,960-token maximum, and the cache would already be 16GiB there. A 500k-token agent context on a model of this shape would need about eight times that for each conversation.

A data center cannot keep every agent’s cache in HBM, so it spills to system memory and then to flash, and comes back as small reads. The Storage-Next guide follows that spill to the SSD and the 512-byte reads it produces, and the High Bandwidth Flash guide covers the flash meant to sit beside the GPU for it. My 4090 run shows a smaller cousin of it on one card. Qwen3.8 27B first put part of itself in system RAM at 65,536 tokens, and from 8,192 to 32,768 tokens the card’s memory in use rose 1.6GiB, as nvidia-smi read it, against 6GiB for the Qwen3-32B estimate over the same range, so Qwen3.8 27B appears to hold far less context per token.

What to watch

No maker publishes a number for the test that matters most, which is whether an agent comes back on time, unprompted, with the work done. A week of timed checks would say more than any launch post.

Figure 8 puts the year on one line. Eleven launches and milestones fell between February and October, five of them since September 8, and the one promise still open has only “later in 2026” for a date.

Three things would move this ranking. Meta’s Confidential VM for Muse, promised for later in 2026, would make it the first of these agents to encrypt a user’s machine with a key the maker does not hold. Duet leaving research preview would show whether Anthropic adds a browser view and errands. Perplexity already publishes what its local mode needs, 24GB or more of memory or VRAM, and Grok Bot and Muse will report their machines when asked, but no maker of a cloud agent publishes the memory and disk each user gets as a spec. The first to do it would let this guide put a firm number on the per-user footprint, which is the figure that decides how much DRAM and flash the agent business buys.