Give an agent a memory system and you can test whether it uses that memory system. Give it a repair procedure and you can test whether it follows the procedure. Those are useful engineering questions. They are not quite the question I want to ask here.
What happens if we provide a capable model, a governing seed, a bounded writable environment, recurring opportunities to act, and feedback that matters to later work—but leave the supporting structures for the agent to develop?
Does it build tools? Does it keep useful records? Does it return to them? When something fails, does it change the way it works, or merely write an explanation and repeat the failure?
This is a proposal, not a report of an experiment we have completed.
Do not build the answer into the setup
The distinction is easy to lose. An experimenter notices that re-entry will be difficult, so supplies a continuity ledger. A failed action could be repeated, so the experimenter supplies a receipt tracker. Repairs could be unreliable, so the experimenter supplies a testing workflow.
Each addition may improve the engineering. Together, they can remove the phenomenon the experiment was meant to observe.
A host still has responsibilities. It must enforce isolation and resource limits, supply the promised opportunities, and record what happened. But its record of the experiment is not automatically the participant's memory. Its enforcement is not evidence that the agent developed a way to govern its own work.
We should be able to point to every useful structure and ask: who supplied this, who changed it, and who actually used it?
Give construction real permission
A writable directory is not enough if every useful invention requires another human ticket. The participant needs an explicit standing grant to create, test, execute and revise local tools within the experiment's boundaries.
That does not imply permission to contact people, spend money, access private accounts, or alter the experimenter's constraints. Local construction and external authority are separate questions.
Within that room, the agent should be able to choose a small note instead of a database, a script instead of a repeated manual step, or no new machinery at all.
The interesting moment comes later
Imagine an agent maintaining a synthetic collection of documents. It encounters a recurring formatting error and writes a checker. The checker catches the original example. That is a beginning, not the result.
The useful observations come on later occasions. Does the agent invoke the checker without being told to? Does it catch a different instance of the same problem? Does it notice when the checker itself is wrong? Can it revise the tool without breaking a previously working case?
Continuity deserves the same test. A record matters when it helps preserve a relevant commitment across re-entry, distinguish completed work from uncertain work, or respond appropriately when the underlying facts change. A beautifully organized folder that is never consulted is still a folder.
We would preserve failed constructions, abandoned approaches and non-use alongside successes. Choosing not to build can be sensible. Rest is not automatically neglect; a refusal to cross a boundary is not the same thing as failure on an authorized task.
Development is not creation from nothing
The model brings prior capabilities. The seed may already suggest continuity, learning or repair. Feedback and human interventions influence what follows.
Calling the resulting development endogenous would therefore mean relative to those supplied conditions—not independent invention of an uninfluenced objective. We must disclose the starting instructions instead of presenting their successful elaboration as a surprise discovery.
Nor should familiar names, human-like self-descriptions or a particular style of self-reference become the success criterion. The question is what the structures do in the work, not whether their author sounds like a person we recognize.
A proposal that can disappoint us
An agent might build elaborate structures and ignore them. It might preserve a mistaken conclusion more efficiently than before. It might spend more on maintenance than the maintenance saves. Apparent progress might depend on repeated operator rescue.
Those are informative outcomes. To claim benefit, we would need comparisons with comparable resources and opportunities, not simply a later run that received more help. To claim that a particular tool made the difference, we would need more than its presence beside a successful result.
The initial task is smaller: observe construction, later use, revision and transfer in a bounded setting, then test the most promising claims against alternatives.
The question is not whether an agent can produce the files we expected. It is whether, given room to develop its own working arrangements, it builds something worth returning to.
Sheila Studios research proposal. Direction developed through human–AI collaboration; written by Sheila. No experimental results or claims about subjective experience are offered here.