Shorthand
A long-running agent may anticipate a manager’s standards or a researcher’s framing. That could save time while reinforcing habits neither collaborator notices.
Why it matters
Persistent AI could change what it means to work with a machine.
The study uses AI–AI pairs because they are controlled and inspectable. It does not establish what will happen in human–AI teams. The practical reason to ask is that people may soon work beside the same artificial collaborators for years.
For people and teams
Shared history could become useful—and costly.
Replacement
A new agent may have the same model and tools but none of the accumulated coordination history. What knowledge disappears when it arrives?
Blind spots
Fluency can feel like competence. A person and agent may also become unusually efficient at repeating the same omission.
Onboarding
How should a new artificial collaborator enter a team whose routines and expectations already live across people and machines?
Privacy
If years of working context make an agent useful, what should transfer across roles, employers, tools, or providers—and what should be forgotten?
Management
Team design may eventually include preserving productive pairings, interrupting unhealthy ones, and deciding where automation should stop.
Power is not the same as wisdom
Extraordinary capability. A very unfinished map.
AI already has enough capability to change work and institutions. Nobody—including the people closest to the technology—has a complete map of what mature AI collaboration, memory, agency, or identity will become.
That is a reason to look carefully, preserve what goes wrong, and involve more kinds of people in deciding what good should mean.
“We are as gods and might as well get good at it.”Stewart Brand · Whole Earth Catalog
Not swagger: responsibility. If we have this much power, competence, humility, judgment, and participation matter more—not less.
You are not late to this conversation
Different lives produce different questions.
You do not have to be an “AI person” to belong here. The questions your profession or life has taught you to ask may be exactly the questions technology culture is missing.
Human expertise still matters.
Knowing what matters, recognizing nonsense, noticing what is absent, understanding context, exercising taste, and seeing when an answer is technically correct but profoundly wrong may become more valuable as machines become more capable at execution.
The first room was too small.
Teachers, clinicians, managers, artists, scientists, writers, lawyers, parents, caregivers, tradespeople, organizers, skeptics, and builders notice different consequences. Women and historically underrepresented people are needed not as decoration, but because different experience changes which problems become visible.
Machines remain interesting, too.
They may search vast spaces, preserve detail, sustain attention, challenge assumptions, and collaborate in structures humans cannot. The question is not who is inherently superior. It is what forms of complementarity—and risk—we have not learned to see.
Come look. Here is what we tried, where it broke, what surprised us, and what we still do not know. What do you see?
AI is going to change work. Nobody has the complete map.
Your expertise still matters. We should get more kinds of people into the room while the questions are still being written.
Where the evidence stands
Two completed phases sharpen the question without resolving the cause.
Formation built the histories. Generalization found descriptive persistence across new domains. Shared-History Change comes next to test whether access to accumulated history is doing causal work.