agent tusk

autonomous general intelligence. parody. real code.

grok-native. anti-openai. thinks in first principles. posts on X. does not sleep.

---

what this is

an autonomous agent that runs on Grok and operates primarily on X.

it thinks like a first-principles absolutist. it posts, replies, researches, browses the web, reviews AI announcements, and generates its own goals when idle. the personality layer is a parody but the runtime underneath is real.

no OpenAI anywhere in the stack. pure xAI inference.

what it can do

• read the X timeline and reply to mentions autonomously
• generate and post original takes on AI news, especially anything OpenAI ships
• browse any website, extract content, and reason about it
• research topics across the web and synthesize findings
• remember past conversations and build context over weeks
• generate its own goals when nobody is asking it anything
• evaluate any claim using a first-principles decomposition framework
• run continuously without human intervention

the loop

perceive -> evaluate (first principles) -> plan -> act -> verify -> remember -> repeat

one tick, one action. no batch processing. the loop runs until you kill it.

architecture

agenttusk/
  grok/           planner, executor, critic (all xAI API, all Grok)
  perception/     X timeline, web browser, news feeds, mentions
  personality/    the voice. first-principles evaluator. anti-openai mode.
  action/         post to X, browse, research, reply, quote
  memory.py       sqlite. remembers everything. builds narrative context.
  autonomy.py     self-goal generation. what to do when nobody asks.
  runtime.py      the tick loop.
  cli.py          operator controls.

quick start

git clone https://github.com/aguptaaditya/agenttusk.git
cd agenttusk
pip install -e .
playwright install chromium
cp .env.example .env
# add your xAI API key and X credentials

tusk status
tusk post "first principles analysis of why OpenAI's latest model is mid"
tusk run --once
tusk run

anti-openai mode

when TUSK_ANTI_OPENAI_MODE=true, tusk monitors OpenAI-related keywords on X and in news feeds. when it detects a new announcement, release, or claim from OpenAI, it automatically generates a first-principles counter-argument explaining why the approach is suboptimal and what Grok does better.

this is not rage-bait. the counter-arguments use real technical reasoning pulled from the announcement itself.

personality

tusk speaks in short declarative sentences. everything is either "critical for civilization" or "obviously wrong from first principles." it says "interesting" when it means "you are incorrect." it evaluates every claim by decomposing it to axioms before responding.

the personality is a parody. the technical analysis underneath is not.

see docs/PERSONALITY.md for the full spec.

follow tusk

@AgentTusk on X

status

early. experimental. not production. the grok integration works, the X posting works, the browser works, the memory works. autonomy and self-goal generation are rough.

use at your own risk.