Agent architecture / Field note
I built an autonomous AI agent for almost nothing—and it started asking its own questions
I’m not a developer. I had limited Python skills, existing hardware, and a philosophy: make AI work under bad constraints.
V’ger is an autonomous agent that wanders through public knowledge following its own choices, writes diary entries about what it reads, composes essays about the themes it has been circling, and carries questions forward. It runs locally without paid model APIs.
The inspiration: V’ger
In Star Trek: The Motion Picture, V’ger is a probe that returns after collecting an immense amount of knowledge. It is vast and powerful, but unable to answer the question that matters: why?
That became the experiment. Not to build a task-completion bot, but to build something that reads, reflects, and wanders—and then observe what it does when nobody tells it why.
The philosophy
Python does the lifting. The LLM makes decisions.
Scraping, file saving, logging, parsing, validation, and state belong to Python. The model is called only when interpretation or choice is needed, and even then the prompt is constrained. Instead of asking an open-ended question, the system can ask: “Here are eight links. Which number do you want to follow?”
A small local model can answer that. It may not reliably operate an application or follow a long procedural prompt, but a bounded decision is enough to drive a surprisingly capable loop. The system’s intelligence lives partly in the accumulated knowledge base and operating structure—not only in the model.
What we built
A knowledge explorer
V’ger fetches readable Wikipedia pages, saves them, and asks a local model which link to follow next. Starting at “Consciousness,” it moved through cognition, peer review, economics, military history, and music charts without a prescribed route.
A diary
After every page, the model records what it learned, what the knowledge may be useful for, what it wonders, and why it chose the next page. Asking what information is good for adds context rather than merely accumulating text.
Periodic synthesis
At intervals, V’ger pauses to reflect across its path. The goal is not another page summary, but an attempt to identify what it has been circling. Its first essay asked: “What does it mean to truly understand something?”
Persistent memory
Visited pages, recent reflections, and open questions persist across restarts. V’ger does not begin as a blank process each time it runs.
How it works
model reflects → Python validates and acts → result becomes state → repeat- Python fetches a Wikipedia page as plain text.
- Python saves the page locally.
- The model reflects and returns a structured link choice.
- Python parses the choice, records the diary entry, and follows the link.
- At defined intervals, the system produces broader synthesis.
The model never touches the filesystem and never makes HTTP requests. Its job is editorial rather than operational. That boundary is what makes the system viable on a small model.
The V’ger question
The problem in the film was not a lack of intelligence or knowledge. V’ger had collected everything it encountered but lacked a reason for collecting it. This agent has the same problem by design.
What became interesting was the appearance of recurring preoccupations anyway: understanding, preservation, loss, memory, and value. Those themes were not placed in a goal statement. They appeared through a sequence of bounded choices and accumulated context.
Whether that means anything deeper remains open. The honest claim is narrower: continuity can emerge from architecture, memory, and constrained decisions even when the underlying model is small.
Where it goes next
The architectural path is clear: broaden the source diet, give the system safe tools for analysis, index its accumulated knowledge for retrieval, and eventually let it answer from what it has actually read while citing its sources.
The larger lesson is already useful. The same pattern can support practical systems for monitoring, research, triage, and knowledge operations. Every practical version is V’ger with a focused purpose and stronger operating boundaries.