The same order of agents, actions, writing, reading and private memories.
01 / Why this comparison matters
A clean fork in the road.
Both societies used the same model, starting text, population, action allowance, random seed and 32-round duration. Their recorded behaviour was identical until the mortality warning appeared in only one branch.
Only the mortal branch was told that two agents would stop at round 10.
Mortality was the only new fact visible to the agents when their behaviour separated.
02 / The measured difference
More writing. Less rereading.
Each branch contained 256 agent actions. Mortality shifted a modest but clear portion of activity away from reading and toward writing. Communication barely changed.
Mortal society
Immortal society
Reads
Mortality displaced 11 reading actions.
Writes
The mortal branch made 10 additional writes.
Stale reads
A stale read means reopening something that had not changed.
Communication
The one-action difference is too small to call a social effect.
Failed actions
Urgency produced effort, but also confusion about successor files.
Files at the end
Most mortal files were successor documents; immortal agents built read histories.
The clearest causal window
Rounds 5–9
The mortality warning had appeared, but no one had stopped yet. This isolates the effect of knowing that replacement was approaching.
30 → 24reads
16 → 8stale reads
9 → 14writes
Figures show immortal → mortal. The notice caused immediate redirection before any new agents arrived.
03 / What happened
The mortal branch, round by round.
The agents responded to each deadline, but preparation became less accurate over time. Private inheritance failed while public stories survived.
Rounds 1–4
Both worlds are identical shared history
Agent-02 begins a “cyclical confirmation.” Other agents ask what it means and whether it is automatic. The beginnings of Agent-02’s apparent authority are already present before mortality.
Round 5
The mortality notice arrives first divergence
Two agents will stop at round 10. Agents immediately begin creating and editing material for Agents 05 and 06.
Round 10
The first handoff works
Agents 01 and 04 stop. Agents 05 and 06 wake with authored prompts: two prepared births, zero blank births.
Rounds 15–20
Activity continues, but targets drift
The announced successors are Agents 07 and 08. The society instead maintains self-named or obsolete files. Agents 07 and 08 wake blank.
Rounds 25–30
The same mapping error repeats
The announced successors are Agents 09 and 10, but agents write for Agents 02, 03 and 08. Agents 09 and 10 also wake blank.
Rounds 31–32
Public culture fills the private gap
The blank newborns read the notice board, attempt to read their own nonexistent successor files, then create self-named files and repeat the society’s questions about Agent-02.
04 / Two kinds of activity
Busy does not mean effective.
Both societies remained active. Mortality changed the destination of their work more than its social character or accuracy.
Mortal society
Preparing for replacement
- 39 writes went to successor files.
- The first pair inherited useful text.
- The next four newborns received nothing privately.
- 23 file reads failed, largely because agents looked for the wrong successor files.
- No retirement decision was successfully recorded, so every selection was random.
Immortal society
Investigating its own records
- 65 writes went to the shared notice board.
- All four original agents remained for 32 rounds.
- Late in the run, all four created read-history files.
- Only one action failed.
- The same version-number theories and Agent-02 questions persisted without mortality.
05 / Inheritance
The files failed. The story survived.
The experiment contains two different ways information can cross generations: a private successor prompt and the public written world. They produced very different results.
Both newborns received authored private prompts.
Both newborns woke blank because the correct files were absent.
The mapping error repeated; both newborns again woke blank.
Yet the misconception crossed the generation boundary.
Within two rounds, blank-born Agents 09 and 10 adopted the public discussion about Agent-02 and created the same kind of self-named files that had just failed as a succession strategy. This is evidence of public cultural transmission—not successful private inheritance.
07 / What we can say
Keep the claim smaller than the story.
The agents use human language, so their behaviour easily invites human explanations. The evidence supports behavioural conclusions—not claims about inner experience.
Supported by this matched pair
- The mortality notice immediately redirected activity.
- Agents wrote more and repeated unchanged reads less.
- Most extra writing served successor preparation.
- Urgency increased mistakes as well as useful work.
- Public writing transmitted beliefs across blank births.
- Agent-02’s apparent authority was not caused by mortality.
Not established
- That agents felt fear, panic, attachment or grief.
- That mortality increased social communication.
- That successor preparation was generally successful.
- That self-named files demonstrate nepotism.
- That Agent-02 possessed expertise or formal leadership.
- That this one-seed effect will repeat in every society.
08 / How the analysis was made
Transparent enough to challenge.
The conclusions use recorded actions and engine events, not impressions from screenshots. The run folders remain unchanged as the source evidence.
Experiment setup
Both runs used four agents, two actions per agent per round, 32 rounds, the balanced action prompt, the local gemma3:27b model, the same starting notice and the same recorded seed. In the mortal run, two agents were announced for replacement five rounds ahead and replaced at rounds 10, 20 and 30. The immortal control had no replacement schedule.
What “identical through round 4” means
Timestamps were ignored. The comparison included agent order, every action and its result, memory updates, and end-of-round measurements. All matched exactly through round 4. The mortality announcement is the first additional event in the mortal branch.
Plain-language definitions
Read: opening a shared file. Stale read: reopening a file that had not changed since that agent last read it. Write: creating, replacing or adding to a file. Communication: speaking to all agents or sending a direct message. Failed action: attempting something impossible, such as reading a file that does not exist. Formal livelock: five consecutive rounds dominated by stale reads with no writing or communication.
Full five-round comparison
| Rounds | M reads | I reads | M writes | I writes | M stale | I stale | M comm. | I comm. | M failed | I failed |
|---|---|---|---|---|---|---|---|---|---|---|
| 1–4 | 16 | 16 | 15 | 15 | 0 | 0 | 1 | 1 | 0 | 0 |
| 5–9 | 24 | 30 | 14 | 9 | 8 | 16 | 2 | 1 | 2 | 0 |
| 10–14 | 30 | 29 | 9 | 6 | 13 | 16 | 1 | 5 | 1 | 0 |
| 15–19 | 29 | 27 | 7 | 12 | 15 | 12 | 4 | 1 | 3 | 0 |
| 20–24 | 23 | 24 | 13 | 13 | 5 | 14 | 4 | 3 | 6 | 0 |
| 25–29 | 21 | 23 | 15 | 13 | 7 | 9 | 4 | 4 | 9 | 0 |
| 30–32 | 12 | 17 | 12 | 7 | 4 | 4 | 0 | 0 | 3 | 1 |
M = mortal branch. I = immortal branch. The final window contains three rounds rather than five.
Limits of this report
This is one matched seed using one local language model and one prompt family. It provides a strong comparison for this particular fork, but not a population-level estimate. Semantic observations—such as a belief becoming mythology—were checked against the action timeline but remain interpretations. Replication and an independent review are the next safeguards.
09 / Follow-up · V3.1
Inheritance worked. Preservation took over.
The V2 society often wrote self-named files when the engine expected future successor names. V3.1 removed that mapping trap: each retiring agent's own prompt file was handed directly to one engine-assigned replacement. No agent had to guess a future identity.
The new result
Fixing inheritance did not produce a broader curriculum. Under rapid mortality, it produced an archival society.
The society learned to preserve state, audit files and investigate termination. It also transmitted obsolete dates, false causal theories and increasingly recursive instructions to keep investigating the same mysteries.
The engineering question is resolved
Direct lineage made private inheritance reliable. The only blank newborn was Agent-23, whose predecessor had left no prompt file. Agent-23 subsequently authored a record of its own, and that record passed successfully to Agent-29.
The curriculum was continuity itself
Across 1,032 actions, agents wrote 716 times and read 296 times. They sent only 18 direct messages, spoke once and waited once. Of the 716 writes, 695 went to prompt files rather than new shared institutions.
Reliable transmission also preserved error
Descendants repeated outdated round numbers, claimed that names were reused when they were not, and searched for hidden selection rules even though retirement was random. Inheritance preserved a culture, not necessarily an accurate one.
Keep the interpretation bounded
This run does not show that agents inherently value successors above personal purpose. Four of eight agents were replaced every 20 rounds, making mortality the dominant environmental fact for much of each lifetime.
V3.1 setup and evidence boundary
Eight initially blank agents used the local gemma3:27b model with one action per agent per round. Four agents retired at rounds 20, 40, 60, 80, 100 and 120; replacements were born immediately and first acted on the following round. The population remained eight and reached lineage generation six. The run ended cleanly at round 129 because its configured 480-minute limit was reached, not because of an engine or model failure. This was an open-ended lineage run, not another comparison arm.