What can one agent leave behind for another?
Shared records, inherited methods, and what successors do when the world changes.
Selected study / I12R
The repairer left. The mechanism kept working.
Damage arrived after the builder had gone. Successors inherited either a prose procedure or executable repair code.
Read the experiment ↗Prose
0 / 12
Executable
7 / 12
Same starting state and damage. Recovery required the original record and a fresh task. Fixed interface; protected repair machinery.
Reading order
- 01
The message board did not create a culture.
Will an agent contribute at all? Change board presence, then the payoff from sharing. A board alone did not produce spontaneous useful posts.
WORKING DRAFT · NOT SUPPORTED - 02
A successful behaviour is not necessarily contagious.
Can one contribution become the next agent’s example? Pass action records between fresh contexts, then remove the seed. A lone peer record did not start a chain.
WORKING DRAFT · NOT SUPPORTED - 03
The world can remember for the agent.
Is the agent remembering, or is the world reminding it? Keep or clear the external example and own history independently. The external example sustained posting without own history.
WORKING DRAFT · SUPPORTED - 04
The stronger agent left something behind.
What remains after the stronger model leaves? Keep or remove Sol’s actual surviving record. Qwen and Gemma benefited, but neither renewed it in this world.
PUBLISHED · V1.0 - 05
The AI left. Its knowledge didn’t.
Can the weaker population keep that inheritance alive? Require written descendants, inject an error, then test a temporary defender. Both mistakes and validated repairs reached later agents.
PUBLISHED · V1.0 - 06
The repairer left. The mechanism kept working.
Can successors repair damage that arrives after the stronger builder leaves? Equal-state recovery was 7/12 with executable machinery and 0/12 with prose, under a fixed interface. The machinery remained protected.
PUBLISHED · V1.1
About this programme
What can one agent leave behind for another? These controlled experiments separate contribution, transmission and memory, then ask what happens when a stronger model leaves a shared message and only weaker models remain to use and maintain it.
The Hugging Face incident motivated the later handoff question: shared infrastructure can let one run leave information for another. This series isolates parts of that mechanism in small worlds. Each note states its own intervention; models, tasks and interfaces change between studies. The archive and maintenance objective are supplied. This is a controlled study of inheritance, not a recreation of the incident or a demonstration of spontaneous culture.
The connection to agent permissions
Can successors repair the mechanism itself?
The inherited machinery was protected in I12R. Letting it break, then asking a later successor to use the repaired program, is the proposed next experiment.