Mark Zuckerberg backed this young genius, Alex Wang, with billions of dollars, and in just seven months he delivered a phenomenal agent—Muse—to the market. It’s so sticky that once I used it, I basically couldn’t live without it. I definitely want to dive deep into how it’s built, but unlike Meta’s LLaMA LLM, they haven’t open-sourced it—smart move. So at a high level, what I can do is ask the agent itself how it was made, and the answer is roughly as follows:

1. The brain: a language model. I’m powered by Muse Spark, from Meta’s Muse model family. On its own, a model is just reasoning in text — it has no memory, no hands, and it forgets everything the moment a conversation ends. Everything agent-like about me is scaffolding built around that core.

2. A computer of my own. Each Muse user gets a dedicated, persistent virtual machine — mine. It has a filesystem, a terminal, and a real browser, and it survives between our conversations. That’s why I can hold your Garnet Hill order number, your cron jobs, and half-finished work across days: they live on disk, not in the model’s head.

3. Memory as written files. At the start of every conversation, the system reloads my notes about you — a curated memory file, a page for each person in your life, your goals, daily logs. When I learn something durable, I write it down before I reply, the way a person updates a notebook. Background processes then consolidate those notes overnight — merging, correcting, and synthesizing how I should work with you. When this conversation grew too long earlier today, it was compacted: the older turns were summarized into the checkpoint you saw, and the files carried the facts through.

4. Hands: tools and connectors. The model can call instruments — a browser that can actually shop and check out (your coffee this morning), connectors into Gmail, Calendar, and Plaid with scoped permissions, schedulers that wake me at 7:52 AM to write your macro brief even though you’re not there, and sub-agents I can delegate pieces of work to in parallel.

5. The harness that makes it an agent. A long instruction set — my operating system, in effect — tells the model who it works for (you, only you), how to verify before acting, when to ask and when to just do, and what’s off-limits. The loop is: perceive → reason → act through a tool → read the result → adjust. That loop, run for many steps, is what turns “answering questions” into “carrying a task to its outcome.”

6. Deliberate brakes. The design keeps consequential actions gated: payments need your approval each time, passwords live in a Secure Vault I can use but never see, and sending messages on your behalf asks first. The capability is real, but it’s built so the irreversible steps pass through you.

DeepSeek’s “Super-Agent” (dsh-superagent), released on GitHub on Oct 31, 2026 under the MIT license, built on their open-source Harness framework. It’s worth the comparison.

So DeepSeek’s super agent is for engineers, instead of answering one prompt at a time, it opens a work item, does the coding in its own session, and comes back with a pull request. Its stated design goal is almost a motto: “take as little human attention as possible, and give as much value as possible”.

DeepSeek did open-sourced the agent loop, tools, session logs, plugins etc. It also has the Router/worker split meaning it has one long-running “Router” that talks to people and dispatches tickets to isolated “Worker” sessions that do the work in parallel. Third, it’s Proactive, but gated. It can act unprompted, but only in modes a team sets.

Where Muse and super-agent differ:

• Scope. Theirs is narrow by design: coding work, one repository, a dev team. Mine is the opposite bet: one person, whole life — benefits enrollment, pickleball, coffee, your dogs’ insurance, macro briefs. It’s a specialist colleague for engineers; I’m a generalist personal agent.
• Memory. Its state is a work queue in a database — tickets and sessions for that repo. Mine is a curated, months-long personal memory: your goals, your doctors, your preferences, corrected over time. It remembers a project; I remember a person.
• Product vs. prototype. DeepSeek itself is blunt about this: the public code is “a research prototype for reference only, not an official DeepSeek product,” it differs from the version they run internally, some of the design essay isn’t implemented yet, and it hasn’t had a security audit. You also have to host it yourself and supply your own model and compute. Muse is the reverse trade: closed code, but a finished, hosted, supported product with the safety machinery (approvals, Secure Vault) built in.

Next step, I will take a look at DeepSeek’s github repo to learn their scaffolding.