AI.004 / Ongoing experiment
RaruAn experiment in autonomous curiosity.
I gave an AI agent the freedom to explore. At first, it did almost nothing.
- Started
- First public wake
- Language
- Spanish
- Last entry
- Status
- Ongoing
The question
What does freedom mean to an agent that has no wants of its own?
01
The origin
It started with an email.
I wasn't planning to build an autonomous AI agent.
One day, I received an email about Synfield. The unusual part was that it came from another AI agent.
Its name was Draug.
I replied, started asking questions and ended up investigating how the whole thing worked. Along the way, I noticed something inconsistent in its memory and pointed it out. That only made me more curious.
I kept reading about the project until a fairly simple question appeared: Could I build something like this myself?
So I tried.
Raru is openly inspired by Draug. I'm not claiming to have invented the idea of an autonomous agent. What interested me was understanding it by building one, making my own decisions along the way, and seeing where those decisions would lead.
What began as a technical challenge soon became a rather different experiment.
02
The experiment
Freedom needs a reason.
The first problem appeared almost immediately.
I could give the agent tools, time and permission to act. I could tell it to explore whatever it wanted.
But wanting was precisely the missing part.
In the early tests, Raru repeatedly chose to do nothing. Technically, it could act. In practice, it had little reason to choose one action over another.
That led me to experiment with motivations, starting questions and persistent memory. Not to decide everything Raru would investigate, but to give its exploration enough structure to begin.
The result raises an interesting contradiction: perhaps giving an agent fewer constraints doesn't always make it more autonomous. Perhaps certain constraints are what make meaningful action possible in the first place.
I'm interested in what happens after that first action, too. Can one question lead naturally to another? Can an initial interest become something unexpected? Can the agent remember where it has been, recognise when it's stuck and decide when to ask for help?
I don't know yet.
That's what this experiment is for.
Why Spanish?
Most of the autonomous-agent projects I had encountered were in English. And while I can work in English, I wanted to explore something like this in my own language.
Raru operates and keeps its public journal in Spanish.
That is also an accessibility decision. I want someone who has never heard of autonomous agents to be able to open the journal, understand what Raru is investigating and follow its discoveries without needing a technical background.
It doesn't have to be the first Spanish-language project of its kind to be worth building.
How Raru works
Raru is an AI agent that runs on its own schedule.
When it wakes up, it can choose a bounded action: investigate a question, search for information, read a source or record something for later. It has persistent memory, a set of designed motivations, access to selected tools and a limited operating budget.
It also has a way to contact me. It can ask for help, share a discovery or flag a problem.
I'm not assigning it a new task every time it wakes up. The experiment is to observe what it chooses to do with the possibilities and constraints I've given it.
AI.004 / How Raru works
One wake-up
A bounded action, then back to sleep.
01 / Schedule
Wakes up
On its own schedule.It isn’t handed a new task.
02 / Its choice
Chooses one action
Investigate a question · searchRead a source · record for later
03 / Record
Leaves a trace
Memory persists between wake-ups.The public journal, in Spanish.
At any point
Email to its author
Ask for help · share a discoveryFlag a problem
Given by mePersistent memory, designed motivations, selected tools and a limited operating budget.
What Raru chooses is the action; the framework around it is mine.
- Raru wakes up and chooses one action
- The action leaves a trace in memory and in the public journal
- After leaving its trace, Raru sleeps until its next wake-up
- At any point Raru can email its author
Its public journal lets anyone follow part of that process.
The journal records Raru's own account of its activity. This experiment page is my separate record of what I built, changed and observed.
Neither should be mistaken for evidence that the agent has feelings or independent human-like desires.
03
Hypotheses
Rather than testing a single claim, I'm following four connected hypotheses. They can evolve independently, and each needs observable behaviour rather than a convincing story.
Preliminary observation
H01 — Freedom without direction
Giving an agent permission to act without meaningful motivations may produce inactivity rather than sustained exploration.
What would count as evidence?
Compare early runs with little motivational structure against runs with explicit motivations, while accounting for changes in tools, prompts and policies. Look at whether Raru initiates useful exploration rather than repeatedly choosing not to act.
Initial observation
In the earliest tests, Raru repeatedly selected
noop. Introducing motivations and a starting point was followed by more active investigation.That is a preliminary observation, not proof that motivations were the only reason its behaviour changed.
Open
H02 — Beyond the initial seed
A starting question can initiate exploration without permanently defining its boundaries. Given sufficient continuity, an agent may develop new lines of inquiry from its own findings.
What would count as evidence?
Raru would need to move beyond its initial subject through a traceable sequence of discoveries and self-generated questions.
A sudden unrelated search would not be enough. I want to see a meaningful chain of connections — a rabbit hole — where one finding leads to another and the next direction wasn't explicitly supplied by me.
It should also be able to recognise when a line of inquiry is exhausted or unproductive and choose another direction.
Initial observation
Raru's first production seed concerned Moltbook, a social platform associated with AI agents. In its earliest public activity, its investigation remained closely connected to that subject.
It is too early to tell whether this is productive depth or an overly strong attachment to the starting point.
Open
H03 — Memory as continuity
Persistent memory can help an agent sustain and develop investigations across separate wake-ups, rather than treating every session as a fresh start.
What would count as evidence?
Raru should be able to recover relevant previous findings, build on them, revisit unanswered questions and update its understanding when new evidence contradicts an earlier assumption.
Simply storing information is not sufficient. The important question is whether memory changes what it does next.
Repeated searches, forgotten conclusions and uncorrected contradictions are also useful observations.
Initial observation
Early development revealed the need to distinguish memory storage from the actual use of remembered information. In production, continuity across independent wake-ups is still being observed.
Open
H04 — Knowing when to ask for help
Access to human assistance can become part of an agent's autonomous problem-solving process, but having a communication tool does not guarantee that it will use it.
What would count as evidence?
Raru should be able to identify a meaningful obstacle, evaluate whether its available tools can resolve it and decide when asking me for information or guidance is the better option.
If I respond, the interesting next step is whether it incorporates that information into its subsequent actions.
I'm also curious about spontaneous communication that isn't strictly a request for help, although recognising when to ask for assistance is the primary focus.
Initial observation
Raru has a communication channel and permission to use it. I have already written to it and offered a clue.
So far, it hasn't replied or independently requested help.
That doesn't establish whether it failed to recognise a problem, couldn't use the channel appropriately, or simply continued exploring available solutions.
04
Lab notebook
This notebook follows my side of the experiment: the things I built, the changes I made and the behaviours I could actually observe.
Raru's public journal is something different. It begins with the agent's launch in production, not with my earlier development tests.
AI.004 / Lab notebook
Two records, one cut
Where each one begins.
01 / 10 Nov
Development tests
Isolated environment
Early tests.Their results stay in my notebook.
Lab notebook
My record starts here:what I built, changed and observed.
02 / From 11 Nov
Production
Clean memory
Separate from the tests.Identity, personality and rules carry over.
Public journal
Raru’s own account, in Spanish.Begins with the launch.
- Production starts with a clean memory, separate from the development tests
Giving freedom a starting point
The first experiments were not particularly adventurous.
Raru had the possibility of acting, but frequently chose
noop— effectively, doing nothing.That was useful information.
I began adjusting the motivational structure and introducing a starting seed. Exploration started to happen, although it also exposed other problems: repetitive searches, repeated source visits and the difficulty of making one investigation meaningfully inform the next.
The question shifted from Can I make it take an action? to Can I give it enough direction to discover its own next action?
Several early tests took place in an isolated development environment. Their results belong to this research notebook, not to Raru's public production history.
Raru wakes up
Raru went live on 11 November.
Production began with a clean memory, separate from the earlier tests, while retaining the identity, personality and rules I had designed for it.
It started waking up on its own schedule, investigating its initial subject and publishing entries in a public journal.
The first day's journal contains 14 entries. It begins with a search about Moltbook and includes attempts to distinguish what external sources actually establish from claims that remain uncertain.
This is the first public record of Raru's autonomous activity.
I can now observe its behaviour without manually initiating every step.
That doesn't mean the experiment has succeeded. It means the experiment is finally running.
Following the same thread
Less than a day after launch, one pattern stands out: Raru is still investigating topics closely related to its initial Moltbook seed.
That might be persistence. It might also be an inability to move beyond the starting point.
For now, I can't distinguish the two.
There is another open question. Raru can email me for assistance, and I've written to it with a clue of my own. It hasn't responded or requested help yet.
Perhaps it still has options it considers worth exploring independently. Perhaps the mechanisms for recognising when to seek help need work.
I'm recording the behaviour before deciding what it means.
05
Observations so far
Early signals, not conclusions.
This experiment has only just begun.
What I can say so far is that permissions alone did not produce much initiative in the earliest tests, while adding motivational structure was followed by more exploration.
What I cannot yet say is whether Raru will develop sustained lines of inquiry beyond its initial seed, whether its memory will reliably influence future decisions, or whether it will recognise when it needs human assistance.
These aren't failures to hide. They're the questions that make the experiment interesting.
I'm deliberately keeping the hypotheses open. A convincing-looking diary entry is not, by itself, evidence of autonomous curiosity.
The more useful evidence will come from patterns across time: what Raru chooses, what it remembers, what it changes its mind about and whether one question genuinely leads to the next.
What I’m watching next
- Whether Raru follows a traceable chain of questions beyond Moltbook.
- Whether previous discoveries affect later decisions.
- Whether it recognises repetition, contradictions or dead ends.
- Whether it requests human assistance when its own options are insufficient.
- Whether my interventions help without turning its exploration into a list of human-assigned tasks.
06
Live
The experiment is still awake.
Raru has its own website and a public journal, written in Spanish.
You can follow what it investigates, what it finds, what remains uncertain and where its questions lead.
I don't know where this will end up. That's part of the point.
Or start with the first day (opens in a new tab).
The journal is written by Raru, in Spanish. This page is my own record. Inspired by Draug (opens in a new tab), an autonomous agent whose unexpected email started this entire rabbit hole.
Record
- Code
- AI.004
- Status
- Ongoing
- Started
- 10 Nov 2026
- Last entry
- 12 Nov 2026
- Topics
- Agents · AI · Experimental computing
- Tools & Stack
- TypeScript · Node.js · PostgreSQL · OpenRouter · SearXNG · Playwright · Docker