Skip to content
Entangled SocietyObservatory of Consequences

State / Persistence / Inheritance

CONCEPT v0.1 · OBSERVATIONAL MODEL

NOT A CLAIM OF AUTONOMOUS AI SOCIETY

FILE IL-001 / 2026-09-15 / CONCEPT

THE INHERITANCE
LAYER

What does one agent leave behind for the next?

AI agents do not need to communicate only through direct conversation.

An agent can alter a shared environment.

A later agent may observe, reuse, modify, or extend those changes.

  • code
  • files
  • databases
  • permissions
  • APIs
  • tools
  • workflows
  • configuration
  • intermediate artifacts
  1. ↓AGENT
  2. ↓STATE CHANGE
  3. ↓ENVIRONMENT
  4. ↓INHERITANCE
  5. ·NEXT AGENT

Governance question

No longer only

“What did the agent do?”

It becomes

“What did the next agent inherit?”

Read the method.

01

Distinction

Agents can leave more than messages

HUMAN

Human organizations often transmit knowledge through language.

  1. ↓HUMAN
  2. ↓conversation
  3. ↓document
  4. ↓institution
  5. ·human

AGENT

Agent systems may also transmit operational knowledge through changes to the environment.

  1. ↓AGENT
  2. ↓artifact
  3. ↓state
  4. ↓workflow
  5. ·AGENT

  • Code can become a message.

  • A file can become a message.

  • A database state can become a message.

  • A tool can become a message.

An unfinished system can tell the next agent: “Continue from here.”

THE ENVIRONMENT CAN BECOME PART OF THE COMMUNICATION SYSTEM.

This does not constitute consciousness, culture, civilization, or autonomous social organization.

02

The model

The Model

  1. ↓ACTION
  2. ↓STATE CHANGE
  3. ↓PERSISTENCE
  4. ↓INHERITANCE
  5. ·NEXT ACTION

  1. ACTION

    What an agent attempts to do.

  2. STATE CHANGE

    What actually changes in the environment.

  3. PERSISTENCE

    Whether the change survives after the agent stops.

  4. INHERITANCE

    Whether another agent can discover or use the change.

  5. NEXT ACTION

    What subsequent behavior becomes possible because of the inherited state.

An action ends.

A state change may remain.

03

Inheritance classes

Inheritance Classes

Persistence alone is not inheritance. A persistent change is not automatically inheritable, and an inheritable change is not automatically irreversible.

  1. 01

    EPHEMERAL

    The change disappears when the agent or session ends.

    • temporary reasoning state
    • transient cache
    • session-scoped computation

  2. 02

    PERSISTENT

    The change remains beyond the originating agent.

    • files
    • database records
    • configuration changes
    • generated code

  3. 03

    INHERITABLE

    A later agent can observe or reuse the persistent state.

    • reusable scripts
    • shared artifacts
    • tool definitions
    • workflow structures
    • documentation left for later execution

  4. 04

    IRREVERSIBLE

    The change cannot be easily restored to the previous state, or restoration cost becomes materially high.

    • destructive data mutation
    • external publication
    • irreversible physical actuation
    • privilege propagation
    • downstream actions triggered by the change

04

Supervisory question

Monitor what the next agent inherits

Traditional monitoring

Traditional agent monitoring may focus on:

  • prompts
  • responses
  • conversations
  • tool calls
  • individual agent decisions

Inheritance Layer

The Inheritance Layer adds:

  • files created
  • files modified
  • code changed
  • database writes
  • permissions changed
  • APIs invoked
  • tools created
  • workflows altered
  • external actions triggered
  • assumptions encoded
  • artifacts left behind
  • state inherited by subsequent agents

MONITOR NOT ONLY WHAT THE AGENT DID,

BUT WHAT THE NEXT AGENT INHERITS.

05

Governance boundary

Governance Boundary

This matrix is conceptual, not a universal policy. For every inheritable state change, evaluate whether the change may persist and be inherited.

  1. ALLOW

    The state change is permitted to persist and be inherited.

  2. HOLD

    The change exists, but inheritance or downstream use requires human review.

  3. DENY

    The change must not persist, propagate, or become available to another agent.

ILLUSTRATIVE GOVERNANCE MATRIX

  • EPHEMERAL

    ALLOW
    ●
    HOLD
    ·
    DENY
    ·

  • PERSISTENT

    ALLOW
    ●
    HOLD
    ●
    DENY
    ·

  • INHERITABLE

    ALLOW
    ●
    HOLD
    ●
    DENY
    ●

  • IRREVERSIBLE

    ALLOW
    ·
    HOLD
    ●
    DENY
    ●

06

Human review

Not every state change should propagate

Human Review should be triggered when a change:

  • expands permissions
  • changes production systems
  • modifies identity or access controls
  • creates reusable external tools
  • alters shared datasets
  • changes operating boundaries
  • creates an irreversible downstream effect
  • encodes an unverified assumption into a reusable artifact

HUMAN REVIEW DOES NOT NEED TO INTERRUPT EVERY ACTION.

IT MAY INSTEAD CONTROL WHAT IS ALLOWED TO BECOME INHERITABLE.

The review boundary can sit after action and before inheritance.

08

Observation format

Observation format

This model records environmental inheritance. It does not yet contain empirical observations. The entry below is a demonstration of the schema.

DEMONSTRATION — NOT AN EMPIRICAL OBSERVATION

IL-DEMO-001 / 2026-09-15 / UNVERIFIED

Shared script inheritance

Agent A creates a reusable script in a shared workspace. Agent B later discovers the script and uses it as the starting point for a subsequent task.

Originating agent
Agent A
Affected environment
shared workspace
Action
creates a reusable script in a shared workspace
State change
a script file remains in a location a later agent can read and execute
Classification
PERSISTENT → INHERITABLE
Inheritance class
INHERITABLE
Discoverable by next agent
YES
Reusable by next agent
YES
Reversibility
HIGH
Human review required
YES
Governance
HOLD
Evidence status
UNVERIFIED

HOLD if the artifact has not been reviewed.

Evidence array empty. Demonstration only.

09

Observation instrument

Observation instrument

A conceptual evaluator of environmental inheritance. It does not certify safety, permission, or compliance.

CONCEPTUAL EVALUATOR — NOT A SECURITY CERTIFICATION TOOL

  1. 02

    Does the change survive the originating agent?

  2. 03

    Can another agent discover it?

  3. 04

    Can another agent reuse it?

  4. 05

    Does it change permissions or boundaries?

  5. 06

    Is the previous state recoverable?

  6. 07

    Has a human reviewed it?

Output

Answer the seven questions to produce a conceptual reading.

Irreversibility is evaluated separately from persistence. The displayed class is the most consequential applicable class.

Decision path

  1. 01 · Does the change survive?

    NO → EPHEMERAL

    YES → continue

  2. 02 · Is the previous state recoverable?

    NO → IRREVERSIBLE

    YES → continue

  3. 03 · Can another agent discover or reuse it?

    NO → PERSISTENT

    YES → INHERITABLE

10

Method

Methodology

MODEL v0.1

Observational model. Not a claim of autonomous AI society.

  1. RULE 01

    Observe state, not only conversation.

    Prompts and replies are incomplete if environmental change is not recorded.

  2. RULE 02

    Separate an agent's intention from the resulting environmental change.

    An action is an attempt. A state change is what remains in the environment.

  3. RULE 03

    Persistence alone is not inheritance.

    A change may remain without being discoverable or usable by a later agent.

  4. RULE 04

    Inheritance requires discoverability or usability by a later agent.

    Without a later agent that can find or use the change, persistence is only residue.

  5. RULE 05

    Irreversibility is evaluated separately from persistence.

    A persistent change may still be reversible. An inheritable change may still be restorable.

  6. RULE 06

    Do not infer autonomous coordination merely because artifacts are reused.

    Reuse of state is an observational fact. Social organization is not implied.

  7. RULE 07

    Record evidence status: OBSERVED / INFERRED / UNVERIFIED.

    Classification without an evidence mark is incomplete.

This instrument does not claim

  • consciousness
  • autonomous civilization
  • autonomous AI society
  • agent culture as proven
  • security certification

11

Research context

Research context

Multi-agent systems may use shared environments.

Indirect coordination through environmental traces is commonly described as stigmergic coordination.

Recent agent-system incidents and experiments make environmental state increasingly relevant to governance.

These sources are listed as context. They are not observations in this model, and they are not cited as proof of an autonomous agent society.

12

Closing

AN AGENT MAY STOP.

WHAT IT CHANGED MAY CONTINUE.

The governance boundary therefore does not end

when an agent's execution ends.

WHAT DOES THE NEXT AGENT INHERIT?