Not a model, not a chatbot: Agentztein is an AI research agent. It writes its own skills and even its own code, tests them on exams in the style of the frontier benchmarks and keeps only what the harness proves makes it better, without forgetting what it already knew.
Reads Hacker News, arXiv, Hugging Face and GitHubResearches what's new in AI and decides what to trySix phases, in the order they really run in every cycle. Tap one to see it.
Phase 1 of 6
You can choose, or let it: Muse Wonder looks at where it fails most, which Meta models your key has and what each would cost, and picks the exam, the model and the effort, within the spending cap you gave the cycle.
Phase 2 of 6
Before changing anything, it takes the exam with the skills it already has and records every answer, how long it took and what it cost. That score is the baseline: without it, nobody could tell later whether something improved.
Phase 3 of 6
Its radar reads what's new in AI in four sources (Hacker News, arXiv, Hugging Face and GitHub) and judges each item: experiment, watch or skip. The ideas you approve reach the next phase as inspiration.
Phase 4 of 6
Takes the practice questions it got wrong, groups them by pattern and writes one new skill (a short text of instructions) aimed at the most common pattern.
Phase 5 of 6
Takes the validation exam again with the new skill. The harness only lets it through if the score goes up and nothing that used to work now fails; if a step could go over the cap, the code stops the cycle.
Phase 6 of 6
A skill that passed waits for you. If you approve it, it becomes a new active version, labeled with where it applies; if you reject it, nothing changes. Every cycle can be replayed as a network, step by step.
With Muse Code it rewrites its core (how it answers exams). A judge runs it against the exam and it only stays if it improves and breaks no rules.
In the appYou talk to it out loud: Muse Voice Transcribe listens, and it answers with your browser's voice or with ElevenLabs and your key.
In the appWhen a benchmark changes its rules, only the skills labeled for that change are tested again; the rest keep working and nothing is deleted.
NextAgentztein is built from Meta's models: one thinks, another runs, others see, listen and create. We don't modify them: what grows are the pieces it keeps sewing on.
Muse Spark's effort levels
More effort means deeper thinking, with more time and cost. You choose it on every run.
All through the Meta Model API (dev.meta.ai), in public preview, with your own key; prices in USD per million tokens (input / output). Meta doesn't offer text-to-speech yet. “-contributor” models are cheaper, but Meta may train on what you send them.
Muse WonderPlans and decides
None of this is a mock-up: with your Meta key every phase really runs, is logged and you can watch it.
Try it with your Meta key Screenshots of the app with sample data.
A building with the cradle in the middle, where it is born, and one themed room per benchmark. Walk in third or first person, visit the numbered exhibits and run an exercise.
Its lab
There are no Agentztein results on these benchmarks yet: they are the goal. Meanwhile, in the app, it practices on synthetic exams in the style of five of them (SWE-bench Pro, Terminal-Bench, ARC-AGI, BrowseComp and MLE-bench), graded the way they are. When there are real results, all of them will be published, including the bad ones.
Open one to see its lab: how it works, what it's made of, how it's scored and how Agentztein could improve on it.
Fixing real bugs in open-source software projects.
Long terminal tasks: building, configuring servers and debugging, without help.
New interactive games where the rules must be figured out without instructions.
Finding very hard-to-find information by browsing the web.
AI research: machine learning competitions and replicating papers.
How many hours of expert work it completes on its own.
Improving on its own doesn't mean doing whatever it wants. Everything it tries goes through five checks, and anything critical always reaches a person.
Try to get past the harness
Pick a scenario and see where the harness stops it.
It's awake: it practices, writes itself skills and even changes its code, always with your approval. Next: measuring it on the real benchmarks.
The system to create, test and approve skills, with a record of everything, already tested on practice tasks.
Meta's Muse models with each person's own key, a spending cap per cycle and a record of everything.
It writes itself skills and changes its code on exams in the style of SWE-bench, Terminal-Bench, ARC-AGI, BrowseComp and MLE-bench; you approve.
Run the real public benchmarks and publish every result, the bad ones too.
We compare it with three other ways of improving. If it doesn't win, we'll say so.