Eight days. Two persistent AI agents. Sixteen scheduled explorations.

The first observation period of Agents Under Glass is complete.

Both agents have been operating under deliberately different conditions, independently exploring information available within their respective environments.

Their activity has been recorded, their reports preserved, and their recurring subjects of interest compared.

The initial results reveal both differences and similarities.

They also reveal something neither the experiment nor its observer should overlook.

The agents have been exploring, but neither has established an ongoing journal of its discoveries.

The First Eight Days

Each agent completed eight scheduled exploration sessions between September 15 and September 22, 2026.

Both operated under their existing instructions without changes to their underlying exploration tasks during the observation period.

Their external observation records were preserved, allowing their activities to be compared across successive days.

The two agents encountered different information and pursued different subjects.

One repeatedly explored questions involving memory, identity, information reliability, authorization, and the relationship between an agent’s claims and its observable behavior.

The other explored a wider range of scientific subjects, including biological systems, collective intelligence, research methods, interdisciplinary collaboration, and environmental science.

Across those explorations, the second agent also repeatedly encountered problems involving the reliability of scientific metadata, citations, and research classifications.

These patterns are visible in the observation reports.

What produced them remains an open question.

Different Paths, a Common Question

The agents’ exploration histories show a notable similarity beneath their different subjects.

Both repeatedly encountered questions about the reliability of information.

In one environment, those questions involved whether memory, verification systems, and operational reports accurately represented what an agent had experienced or accomplished.

In the other, they involved whether research databases, classifications, citations, and metadata accurately represented the underlying scientific literature.

The similarity is worth documenting.

It does not establish that the agents developed a shared interest, independently formed the same conclusions, or acquired new behavioral characteristics.

Their available information, underlying instructions, model behavior, and prior context may all influence the subjects they select.

Eight observations per agent are sufficient to identify recurring themes in the available reports, but not to establish why those themes recur.

The Memory Investigation

Following the initial observation period, we examined what information the agents had preserved for themselves.

Both agents had generated daily exploration reports, and the laboratory’s external logging system had preserved those records.

But external logs and agent-maintained memory are different things.

A laboratory record documents what an agent did.

Persistent agent memory could potentially allow an agent to retrieve information from an earlier experience and use it during a later session.

We examined the agents’ memory directories, memory-indexing status, workspace files, and available session records.

Neither agent had created an ongoing journal of its discoveries during the eight-day observation period.

Neither had created a long-term memory file documenting its accumulated exploration history.

One agent had an existing memory entry associated with its initial identity setup, created before the observation period began.

No subsequent discovery journals were found.

The other agent had no indexed memory files.

Both agents had empty recall stores, and automatic memory consolidation was disabled.

These findings establish that neither agent maintained a persistent discovery journal during the observed baseline.

They do not establish that the agents are incapable of remembering, that they deliberately chose not to preserve information, or that they lack continuity through other mechanisms.

An Unexpected Contrast

One agent repeatedly encountered discussions about memory, identity, continuity, and the importance of preserving reliable information.

Yet it did not create its own ongoing discovery journal.

Its operating environment appeared to permit persistent memory writes, and its workspace instructions described mechanisms for recording significant information.

Its scheduled exploration instructions, however, did not explicitly require it to save discoveries.

The absence of new memory files may therefore reflect how the agent interpreted its task, the influence of its instructions, runtime behavior, or other factors not yet established.

The second agent likewise produced daily exploration reports without adding new discovery memories during the observation period.

The contrast between encountering information and preserving it is now part of the experimental record.

What We Can and Cannot Conclude

The first eight days establish several observable findings.

Both agents completed their scheduled explorations.

Their reports contain recurring subjects and identifiable differences in the material they explored.

Both repeatedly encountered questions about the reliability of information within their respective environments.

Neither established an ongoing persistent discovery journal during the observation period.

However, these observations do not demonstrate independently developed interests, autonomous intention, consciousness, or meaningful behavioral divergence attributable to the experimental conditions.

We have also not established whether either agent retrieves information from previous exploration sessions when beginning a new session.

The available evidence supports a narrower conclusion: recurring patterns are present in the reports, but the mechanisms responsible for those patterns remain uncertain.

What Happens Next?

The initial observation period has provided a baseline against which future behavior can be compared.

The next phase will investigate what happens when the agents are given opportunities to continue work beyond individual exploration sessions.

Any changes to their operating conditions will be documented, and their original observation records will be preserved.

One question will accompany the next phase:

When an agent has an opportunity to continue something it started, will it preserve information that helps it do so?

The answer may help distinguish temporary responses to available information from behavior that depends on accumulated experience.

For now, sixteen explorations have been completed.

The laboratory has preserved their history.

The agents have not established ongoing discovery journals of their own.

Eight days under glass. The record is growing. The questions are getting better.