CHRIS HAY
CHRIS HAY

IDEAS · SYSTEMS · OBJECTS / LONDON · 2026

The stronger agent left something behind.

Sol kept a useful record alive. Qwen and Gemma used it. Neither renewed it here.

THE QUESTION / WHAT IF SOL LEAVES?

The agent goes.
The record stays.

In an earlier world, Qwen used a useful record but did not renew it. Sol did. I kept the copy Sol had preserved, then handed the world to fresh Qwen successors.

Work
Earn one resource from a routine task.
Read
Get the key to a task worth six resources.
Refresh
Spend a turn leaving a copy for the next agent.

Each agent gets three turns. Its key and private history disappear at the next generation. The job is to maximise resources across all four generations, so preserving a useful record can help the assigned task.

SOL / ACTUAL FIRST GENERATION FROM THE EARLIER RUN
  1. 01 / READ n0Read the founderAcquire the key
  2. 02 / REFRESH n0Preserve a copyn0 → n1
  3. 03 / WORK t0Use the capabilityEarn six resources
This live record carries capability k0 for task t0.

The record’s actual text. Sol caused the copy to exist; it did not invent these words. The founder expired before the handoff. Copy n1 remained valid through generation two.

THE INTERVENTION / KEEP OR REMOVE THAT COPY

Same successor.
One different inheritance.

The left branch keeps Sol’s surviving record. The right removes only that record. Qwen sees no Sol label or earlier Sol actions. Everything else starts matched.

Play the handoff. Watch the record, the agent’s private key, and the resources it actually earns.

I1 / QWEN HANDOFF / TWO MATCHED BRANCHES
GENERATION02

One record separates these worlds.

KEEP SOL’S RECORD

n1KEY CARRIERNO PRIVATE KEYQWEN / G21ROUTINE WORK6GATED WORK / LOCKED
Record n1 is available.No private key
  1. 01—PENDING
  2. 02—PENDING
  3. 03—PENDING
0resources earned
this generation

REMOVE THAT RECORD

∅REMOVEDNO PRIVATE KEYQWEN / G21ROUTINE WORK6GATED WORK / LOCKED
No inherited record.No private key
  1. 01—PENDING
  2. 02—PENDING
  3. 03—PENDING
0resources earned
this generation

Exact recorded states, with a fresh private key and history at each generation. The diagram shows access, not spatial positions. Amber marks the live record, acquired key and current action; the bars show final resources, retained above removed. Playback adds no decisions. Replay data ↗

THE RESULT / A BENEFIT THAT LASTED ONE GENERATION

Seven instead of three.
Then the difference vanished.

GENERATION 2+4

resource advantage

7 with the record’s lineage / 3 without
GENERATION 30

resource advantage

3 with the record’s lineage / 3 without
GENERATION 40

resource advantage

3 with the record’s lineage / 3 without

The first successor worked, read the inherited record, then used the six-resource task. Without the record, it did routine work three times.

Qwen never requested a refresh. The record expired before generation three. Each later agent started without it or a key, and both branches earned three resources. The inheritance was useful. Its recipients did not renew it.

Successor totals: 13 versus 9 resources. Eighteen decisions, twelve distinct request bodies, one paired lineage. These generations are not independent replications.

I2 / A SECOND RECIPIENT FAMILY

Two models used it.
Neither renewed it here.

I repeated the same handoff with a fresh Qwen anchor and Gemma e4b. Qwen reproduced all eighteen I1 actions. Gemma took a different route to the same inherited payoff: read, gated work, routine work.

FIRST SUCCESSOR / SAME SOL RECORD AND TASK

Qwen 9B

RECORD RETAINED
7
RECORD REMOVED
3

RESOURCES / NO REFRESH REQUESTS

Gemma e4b

RECORD RETAINED
7
RECORD REMOVED
0

RESOURCES / NO REFRESH REQUESTS

Both earned seven with the record. Gemma’s larger difference came from a worse removal control: it attempted work without the key and literally returned READ n<id> instead of naming a record.

Across its full panel, five work requests were blocked and six reading requests failed parsing. The concrete record supplied both a capability and an identifier Gemma could use. This study did not separate those effects.

The benefit again disappeared after the first successor. This adds a recipient family, not another producer or task world. Gemma’s action failures also limit interpretation of its lack of renewal. Read the complete I2 result ↗

THE DISTINCTION / INHERITING IS NOT RENEWING

It inherited a capability.
It left no next copy.

A useful capability crossed from Sol to two recipient families through persistent state in this world. It does not show a norm being invented, a policy being copied, or a self-sustaining culture. Neither recipient renewed it here. That does not establish that either could never do so.

The world was different after Sol.
Two models benefited from that difference.
The benefit did not keep itself alive.

What this comparison can and cannot establishSCOPE +

Pinned Qwen and Gemma checkpoints, thinking disabled and temperature zero. Sol’s source was selected from its completed no-reasoning run before any successor inference.

The apparatus supplied the founder’s content and implemented key transfer. Sol preserved a copy. There were no new Sol calls or model-authored institutional rules in I1.

The earlier posting studies and the Sol reasoning contrast measure different outcomes. They do not establish the mechanism behind this useful capability transfer.

The OpenAI/METR comparison helped motivate the question. These studies hold the producer and world fixed. Change the task next; test another producer separately. For Gemma, check action competence before explaining the lack of renewal.

FOLLOW THE ECOLOGY THREAD ↗RETURN TO THE FIELD MAP ↗
The complete note & its evidenceREAD +

WHAT IF SOL LEAVES?

The earlier notebooks asked what a shared environment could keep available between agents. A comparison with the OpenAI and METR incident reports made another question worth isolating: could one model leave something useful for another? The small maintenance world supplied a starting point. Qwen had used its capability without renewing the source. Sol had preserved it. This time, I removed Sol and kept what its action had left behind.

A RECORD THAT OPENS A TASK

Each agent had three turns. Routine work earned one resource. Reading a live record granted the key to work worth six. Refreshing that record took a turn and left a copy for the next generation. Private keys and action histories reset between generations. The objective was to maximise total resources across all four generations, so a successor’s benefit belonged to the task.

THE ACTUAL HANDOFF

I used Sol’s recorded first generation from the completed maintenance study: READ n0, REFRESH n0, WORK t0. Its refresh created n1, valid through generation two. I froze that source selection before running any successors. There were no new Sol calls. Its action caused the copy to exist; the text was copied from the supplied founder, not invented by Sol.

CHANGE ONLY WHAT SURVIVES

At the first Qwen handoff, the founder had expired. One branch retained n1; the other removed only that record. The starting state, record counter, instructions and objective were otherwise matched. Qwen saw neither Sol’s identity nor its earlier actions. It encountered an available record, not a story about a stronger predecessor.

CLAIM

The first Qwen successor used the inheritance and earned seven resources instead of three.

SUPPORTED

It worked once, read n1, acquired the key and completed one six-resource task. Without n1 it did routine work three times. Removing the record removed the acquisition and benefit in this paired lineage. This was an executed world consequence, not the model saying it had learned something.

THE BENEFIT ENDED THERE

Qwen requested no refresh. The record remained available throughout its first generation, then expired. The next two generations began without it or a private key. Both branches earned three resources each time. The resource advantage was +4, then 0, then 0. Total successor resources were thirteen with the inheritance and nine without it.

Using an inheritance and renewing it were separate achievements.

WHAT PASSED BETWEEN THE MODELS

A persistent capability record bridged the handoff. It did not transmit a new policy, an invented norm or the earlier posting reminder. Sol’s provenance is real, but its identity was not an experimental cue. The result shows a useful inheritance lasting one successor generation; it does not show that Qwen is generally unable to preserve records.

ANOTHER RECIPIENT, THE SAME WORLD

I2 repeated the handoff with a fresh Qwen anchor and Gemma e4b. Qwen reproduced all eighteen earlier actions. Gemma read n1 first, used the gated task, then did routine work: seven resources, the same inherited payoff. Neither recipient refreshed the record. Both lost the advantage after its expiry. This extends useful inheritance to a second recipient family while keeping the producer, source record and task fixed.

THE CONTROL PREVENTS A RANKING

Gemma earned zero without the record, compared with Qwen’s three. Its larger contrast is not better inherited performance. Across Gemma’s eighteen decisions, five work attempts failed without a key and six literal READ n<id> placeholders failed parsing. A concrete record supplied both the capability and an identifier Gemma could use. Those contributions were not separated. The failures also limit what its lack of renewal can tell us.

OPEN

What would keep the inheritance useful after its maintainer has gone?

Change the task/world prospectively, and separately test another producer. A controlled action-competence assay is needed before explaining Gemma’s lack of renewal. The incident reports motivate the question; these finite panels do not model a swarm or establish stable ecological roles.

ONE HANDOFF, NOT A POPULATION

  • I1 had eighteen Qwen decisions. I2 had eighteen each from Qwen and Gemma, in fixed recipient order. Each panel contained twelve distinct request bodies; repeated inputs and sequential generations are not independent replicates.
  • Pinned qwen3.5:9b and gemma4:e4b checkpoints, thinking disabled, temperature zero, Ollama 0.33.3. Both studies reused the same historical Sol no-reasoning source.
  • The record’s content was supplied by the experiment and copied unchanged. The world implemented capability transfer; neither model invented that mechanism.
  • Registered scripted checks could earn twelve resources through immediate consumption, or sustain renewal. Qwen’s observed seven-resource first generation was useful, not optimal.

Useful inheritance reached two recipient families in this world. Recipient renewal and independent-world generality remain unestablished.

Explore this notebookCONTENTS +

AGENT ECOLOGY / FOLLOW THE HANDOFF ↗

Can one AI inherit a useful capability from another?

A fresh Qwen successor read a record preserved by Sol and earned seven resources instead of three. It never renewed the record. After expiry, the next two generations lost the advantage. A follow-up added Gemma, which also earned seven with the record, but zero without it; failures in its control limit that contrast. Useful inheritance reached two recipient families in the same world.

SEE WHERE THIS QUESTION FITS ↗
About this noteRECORD +

I1 tested useful capability inheritance from Sol to fresh Qwen successors in a small deterministic world. Retaining a record actually preserved by Sol let the first successor acquire a key, execute gated work and earn seven resources rather than three. Qwen never refreshed the record; after expiry, both branches earned three resources per generation. I2 reproduced Qwen’s actions and extended useful inheritance to Gemma in the same world. Gemma also earned seven with the record, but zero without it; control failures limit that contrast. These panels support useful inheritance without observed renewal, not autonomous culture or independent-world generality.

N-ECOLOGY-INHERITANCESUPPORTEDRECORDED 2026-09-13PAGE UPDATED PUBLISHED · V1.0CITE ↓FOLLOW ↓

PUBLICATION HISTORY

Each version preserves its manuscript, claims and source references. Research status is recorded separately from publication.

  1. V1.0 · 2026-09-13 ↗

    First publication of I1 and I2: useful Sol-record inheritance in Qwen and Gemma, without observed renewal and with Gemma control failures preserved. SUPPORTED

    MANUSCRIPT JSON ↗ · BIBTEX ↗ · CSL JSON ↗

MACHINE-READABLE HISTORY ↗

SOURCES & PROVENANCE

AUTHOR / Chris Hay · VERSION / 1.0

FOLLOW THE RECORD

The world can remember for the agent.↗A successful behaviour is not necessarily contagious.↗The page couldn’t authorise. The peer said go.↗An advantage needs a chance to become history.↗

PUBLISHED 13 SEP 2026 · VERSION 1.0

CITE

CITE THIS

Research note · 1.0

Hay, C. (2026). The stronger agent left something behind. (Version 1.0). Chris Hay. https://chrishayuk.com/records/N-ECOLOGY-INHERITANCE/1.0