The Approach
Agents Under Glass is an observational experiment built around comparison, persistence, and record integrity.
Multiple persistent AI agents operate under deliberately different conditions. Their behavior and development over time are compared against a privately maintained experimental record.
Some details of the design are intentionally withheld while the experiment is active. This is not secrecy for its own sake. Public disclosure of subject identities, experimental assignments, active hypotheses, or certain environmental conditions could influence the systems being observed.
The challenge is determining which differences are observations, which have plausible technical explanations, and which deserve further investigation.
Comparative Design
The agents begin from closely matched technical foundations, with differences introduced deliberately rather than casually.
Where practical, their underlying systems are kept technically comparable so that behavioral changes can be considered alongside differences in experience and environment.
The precise assignment of experimental conditions, subject identities, and certain environmental details are withheld from the public record while observation is active.
Perfect experimental isolation is neither assumed nor claimed. When differences in conditions are discovered, they become part of the research record.
The Observation Record
Agent behavior is documented as it occurs rather than reconstructed later from memory.
Significant observations are preserved with supporting source material and integrity references where appropriate, allowing later analysis to be compared against the record available at the time.
Supporting records are maintained privately so that important observations can later be examined against the evidence available at the time.
Public Field Notes are derived from that record but may omit or delay details when disclosure could influence the systems being observed.
Observation, Interpretation, and Speculation
The record distinguishes between three levels of reporting.
Observation describes what directly occurred in the available record.
Interpretation considers plausible explanations for that observation based on the surrounding context and known system behavior.
Speculation explores possibilities that are not established by the available evidence.
These categories matter because unusual AI behavior can have ordinary explanations. Prompt context, memory retrieval, model variability, configuration changes, software behavior, or coincidence may produce something that initially appears significant.
Interesting is not the same as unexplained.
Change Control
Changes to the experimental environment are treated as part of the experiment.
Software updates, model changes, configuration adjustments, memory changes, new tools, and other interventions can affect agent behavior. Material changes are therefore documented when they occur.
Whenever practical, comparable technical conditions are maintained across the experiment. When meaningful differences arise, they are recorded rather than ignored.
The objective is not to create perfectly identical systems forever. It is to know when, where, and why the experimental conditions changed.
Reproducibility and Limits
Agents Under Glass is not a controlled laboratory study in the academic sense, and it should not be mistaken for one.
The experiment uses real AI systems, software, hardware, network services, and external environments that may change over time. Some variables can be controlled. Others cannot.
Supporting technical records are preserved privately so that significant observations can be examined against the evidence available at the time.
The purpose is not to manufacture certainty.
It is to leave a trail.
Contamination Control
Public documentation creates an unusual problem for an experiment involving systems capable of accessing public information.
Information published here could potentially become part of the environment being observed.
For that reason, subject identities, experimental assignments, active hypotheses, and other identifying details may be withheld, generalized, or published after a delay.
If evidence suggests that a subject has independently encountered this site or information derived from it, that discovery will itself be treated as an experimental event and documented in the private research record.
The objective is simple: observe without unnecessarily becoming part of what is being observed.