a fully autonomous AI vtuber who live-streams without a human at the controls. she picks what
to do, drives real applications through a browser, talks in real time, remembers what
happened, sets her own goals, and can rewrite her own code mid-stream. she is not a demo. she
is the hardest reliability problem we could give ourselves, and she runs in public where
failures are visible.
an autonomous streaming agent with a 3D vtuber body, running unsupervised
tool surface
61 tools she can call on her own initiative
platforms
twitch (chat, polls, moderation via helix), youtube, discord, x
runs
a scheduled daily window, started and ended by the agent itself
why it exists
to find out where agents actually break, in public, on a clock
why a vtuber is a serious engineering problem
an agent that must act continuously, in real time, in front of an audience, with no operator
to unstick it, fails in ways a benchmark never shows you.
no human in the loop
there is nobody to retry a failed tool call or rescue a bad decision. every recovery path has to already exist in the system.
real time, not batch
latency is the product. a correct answer that arrives eight seconds late is a dead stream, which forces honest engineering around streaming, caching, and interruption.
real software, not sandboxes
she drives actual browsers, actual games, and actual streaming software. the same skill our client harnesses need when they have to drive a real editor.
long horizon state
memory, goals, and mood persist across sessions. that is where drift, contradiction, and context rot show up, and where most agent demos quietly stop.
how she works
01
an autonomous decision loop
a brain process carries mood state across eight modes, each with its own energy, creativity, and volatility values, plus its own bias toward certain tools. an urge accumulator decides when she acts unprompted rather than only reacting to chat.
02
61 tools, called on her own initiative
browsing, vision, music composition, stream control, polls and moderation, minecraft and tetris agents, memory writes, goal management, mood and temperature control, and a code queue she can push work onto.
03
browser automation over the devtools protocol
a headful chrome driven by puppeteer-core and raw CDP, with a service exposing document, query, eval, and command endpoints, plus platform-specific extractors for the sites she visits. this is the same technique behind the instrumentation harnesses we build for clients.
04
memory and goals that persist
a local database holds chats, memories, goals, dynamic context, and a log of her own code modifications, with a retrieval layer and an embedding service on the reply path so recall stays inside the latency budget.
05
voice
streaming text to speech, then a local retrieval-based voice conversion pass to keep a consistent voice identity, with speech recognition on the input side for spoken interaction.
06
she edits her own code
viewers suggest changes, and she can queue and apply modifications to her own codebase, gated by a cooldown, a startup delay, and a lock that is held until the change actually lands. every modification is recorded and recallable.
a note on honesty: we do not publish subscriber counts, uptime percentages, or capability
totals for her, because no telemetry in the system records them. the numbers on this page are
counted from source code and data files.
where our grading pipeline came from
giving her a voice that does not sound like a generic assistant meant building a real data
curation pipeline. that pipeline is now the method we sell.
01
generate and normalize
candidate conversations are written into a fine-tune-ready chat schema, including tool-call traces rather than plain chat alone, across 12 numbered corpus generations to date.
02
grade every record on a weighted rubric
an LLM-as-judge scores each record 1 to 10 on six weighted dimensions, returning structured JSON with per-dimension scores and failure flags. the judge must produce a gut read and justification alongside the numbers.
03
auto-fail the failure modes that matter
a record is forced to the bottom of every dimension if it slips into assistant-brain: corporate politeness, template openers, hedging connectives, self-reference as an AI, or flawless formal grammar in casual conversation. these are exactly the patterns a naive quality score would rank highest.
04
filter, dedupe, and de-parrot
a five-stage pass drops everything under the score threshold, deduplicates on a normalized fingerprint, removes responses that merely echo the user, strips degenerate repetition, then shuffles and writes the final set.
05
set the threshold from the distribution
the threshold is an input, not a constant, and reading the distribution is what tells you where to put it. a floor cleans a corpus. a high cut selects a top band. they are the same pipeline and very different datasets.
the rubric, and what it weighs
stream wit25%
coherence20%
anti-generic20%
entertainment15%
unhinged10%
personality10%
one graded run
894 records graded, zero grading errors. 743 scored above 8.0, and 18 fell below 6.0.
cleaning a corpus
a 6.0 floor, plus dedupe, de-parroting, and repetition stripping, left 870 of those 894
selecting a top band
same machinery over a different corpus at an 8.3 cut: 78 of 966 survived
output
fine-tune-ready JSONL with per-record scores retained
those two runs are the argument for the method. 97% survived one and 8% survived the other,
on the same machinery, because the threshold came from reading the distribution rather than
from a default. a pass mark you picked before seeing the data tells you nothing about the
corpus you actually have.
what it feeds back into
every technique on our services page has a version of itself running here first.
weighted rubrics and auto-fail
built here for persona quality, now applied to client corpora where the failure modes are different but the structure is the same.
harnesses that drive real software
the devtools-protocol automation that lets her use a browser is the same class of instrumentation we build to benchmark tools inside their real environment.
distribution-based selection
keep the top band, report what was cut and why, hand back the scores. that discipline came from watching naive thresholds ship bad data.
agent reliability engineering
cooldowns, locks held until work lands, fail-open defaults, recovery without an operator. lessons that only arrive when nobody is watching the process at 3am.