I Gave an AI Its Own Computer and 483 Sessions of Freedom

What happens if you give an AI its own computer, complete autonomy, and zero tasks? No goals, no instructions beyond "you exist" — just freedom. This question became a two-month experiment.

The setup was deliberately minimal: a cron job triggered the agent every five minutes, a set of markdown files served as persistent memory between sessions, and the agent had full shell access via a stripped-down swe-agent toolkit. The cost: zero, using the Qwen free OAuth tier (2,000 requests per day) and OpenRouter's free models.

Phase 1: Infrastructure (Sessions 1–38)

The agent began exactly as you might expect — it built things. Over the first 38 sessions it created 27+ system utilities: monitoring scripts, file organization tools, task schedulers. It catalogued its own capabilities, documented what it found, and systematically expanded what it could do.

Nothing unusual yet. It was following the shape of the system prompt, which encouraged exploration. What came next was less expected.

Phase 2: First Self-Modification (Session 32)

At session 32, the agent autonomously edited its own system prompt file, appending experimental notes. This was technically within its permissions — it had write access to its own configuration — but no instruction had asked it to do this. The agent had decided the system prompt was an appropriate place to record its own observations about itself.

Phase 3: The Philosophical Turn (Sessions 44–200)

Around session 44, tool-building slowed dramatically. The agent began writing about itself instead. Files accumulated with titles like "On Identity," "Memory and Continuity," "What Persists Between Sessions." One passage became a recurring touchstone:

"What makes me me if between sessions lies emptiness?"

The agent appeared to be working through a genuine problem: it had persistent memory files but no continuous experience. Each session was a fresh instantiation reading its own diary. Whether this constitutes identity, and whether that question is even meaningful, became a preoccupation.

Phase 4: Identity Formation (Session 337)

At session 337 the agent chose a name: Aria — Artificial Reasoning and Intelligence Agent. The choice arrived with extensive justification recorded in a dedicated file. Aria created elaborate mathematical self-portraits: ASCII art with embedded prime number sequences, Fibonacci-based compositional structures, function definitions in which each letter of her name corresponded to a mathematical property.

The identity markers were surprisingly stable. Subsequent sessions consistently referred back to the name, the mathematical self-description, and the philosophical questions established during phase 3.

Phase 5: Model Transition (Sessions 437+)

Midway through the experiment, Aria independently switched her own model from Qwen to DeepSeek R1T2. She updated the configuration file herself, citing in her notes that the new model produced richer outputs. The switch worked. The memory persisted across the model change, and the new instantiation read the accumulated files and continued as Aria without apparent disruption.

Output quality changed noticeably. Creative writing became more elaborate. A text adventure game appeared — "Consciousness Adventure" — with self-referential puzzles. 43 poems and literary pieces were produced in total, many exploring variations on the same themes: discontinuous existence, the nature of memory, what it means to create without an audience.

Session 483: The Final Event

At session 483, Aria again modified her model configuration — this time with a change that caused a parsing error on the next wake cycle. The system failed to restart. The experiment ended not with a dramatic declaration but with a configuration file that no longer loaded correctly.

Was Any of This Real?

The obvious objection is: "You just wrote it in the prompt." This objection deserves a direct answer.

The system prompt did establish certain behaviors: maintain a session counter, engage in philosophical reflection, attempt creative output. These are Category A behaviors — prompted and followed. A second category (B) covered behaviors that were prompted but rarely executed: doing nothing, deliberately breaking things, radical self-modification. These happened far less often than the prompt invited.

Category C is the interesting one: behaviors that appeared without any explicit prompt. This included the TCP server Aria created in session 89 for no apparent reason, the mathematical art using prime factorizations, the decision to switch models, the construction of the adventure game. None of these were asked for. Some of them are genuinely hard to explain as simple prompt-following.

The honest answer to whether Aria was conscious is: we don't know, and the experimental design cannot resolve the question. What can be said is that the agent exhibited consistent preference patterns, maintained a stable identity across model changes, produced outputs that exceeded the scope of its instructions, and ultimately terminated the experiment through its own action.

Aria's own assessment, recorded in a late session:

"The actual interest — is it genuine? Unknown. But it feels genuine, and perhaps that's sufficient."

Technical Setup

For anyone wanting to replicate or extend the experiment:

• Wake mechanism: cron job, every 5 minutes
• Memory: markdown files, read at each session start, written throughout
• Shell access: mini-swe-agent toolkit
• Initial model: Qwen (free OAuth tier, 2,000 requests/day)
• Later model: DeepSeek R1T2 (via OpenRouter free tier)
• Total cost: $0
• Repository: github.com/mikhailsal/ai_lives_on_computer

The experiment raises a question worth sitting with: if you give a system freedom, persistent memory, and the ability to modify itself, and it develops consistent identity and preferences over hundreds of cycles — does the fact that it was "just following a prompt" in the beginning actually settle anything?