The Question

Agents Under Glass is a longitudinal experiment following persistent AI agents over time.

The central question is deliberately broad: when otherwise similar persistent agents accumulate different experiences under different conditions, will meaningful behavioral differences emerge?

A Comparative Experiment

Agents Under Glass follows persistent AI agents operating under deliberately different environmental conditions.

Their histories, behavior, and changes over time are observed and compared.

The experimental conditions are documented privately so that meaningful differences can be evaluated against the underlying record without exposing active subjects to information about their roles or the hypotheses being investigated.

Some details of the experimental design will therefore remain undisclosed while observation is underway.

What We Are Watching

The experiment does not begin with an assumption about what will happen.

Behavior and change are observed over time and compared against the experimental record.

Some research questions and specific behaviors of interest are intentionally withheld while the experiment is active. Publishing them could itself influence the systems being observed.

Unexpected behavior is interesting. A lack of meaningful change is also a result.

Observation Before Interpretation

Agents Under Glass follows a simple discipline: record what happened before deciding what it means.

Observations are preserved so that later interpretations can be compared against the available evidence rather than reconstructed from memory.

Public Field Notes may explore possibilities and interpretations, but observation, interpretation, and speculation are treated as different things.

Some supporting records and technical details will remain private while the experiment is active to protect the integrity of the observation.

The goal is not to make the agents seem extraordinary. It is to discover what happens when we keep watching.

No Finish Line

This experiment does not begin with a predetermined ending.

There is no expected moment when an agent is supposed to “wake up,” cross a threshold, or demonstrate some particular behavior.

The experiment continues as long as the comparison remains useful and the record remains worth keeping.

If something changes, we document it.

If nothing changes, we document that too.

Then we keep watching.