The gaps start before peer review
AI has made it easier to explore ideas, write code, and run experiments, and autonomous research or recursive self-improvement could accelerate that process further. Yet the debate over research quality still centers on peer review and publication incentives. Both matter, but the work takes shape long before a reviewer sees it.
We still work as if the paper were the project, leaving the evidence, failed attempts, and decisions behind it scattered across tools and people. By the time a coauthor receives the draft or a PI sees an update, the reasoning they need to question the direction may already be missing.
As more work passes through agents, this gap can grow further, and even the person producing a result may have fewer opportunities to understand how it was reached.
Imagine, you ask a collaborator to help with the manuscript, check the references, and write the related work section. Later, you send your PI the final experimental analysis and the paper for comments. Your data, preprocessing code, and notes explaining decisions, such as why you excluded a batch, stay with you. Everyone is contributing to the same project, but what each person and their AI can contribute depends on the context available to them.
| People / their AI | Literature | Lab notes | Data | Preprocessing | Experiments | LaTeX |
|---|---|---|---|---|---|---|
ALResearcherAI with this context | Not in context | lab-notes.md Exclusion noteBatch B excluded. Check generalization. | samples.csv idbatchscore041A0.82042A0.79043B0.31044B0.28 | prepare.py df = load()df = df[ batch != "B"]fit(df) | analysis.ipynb Reported improvement+12% | manuscript.tex \section{Results}Results Our method improves performance by 12%. |
MRCollaboratorAI with this context | references.pdf Evaluation methods[1] [2] [3] | Not in context | Not in context | Not in context | Not in context | manuscript.tex \section{Results}Results Our method improves performance by 12%. |
SCPIAI with this context | Not in context | Not in context | Not in context | Not in context | analysis.ipynb Reported improvement+12% | manuscript.tex \section{Results}Results Our method improves performance by 12%. |
If we want AI to improve how research is done and help collaborators make meaningful contributions, people and their agents need access to the context behind each result: how it was produced, which decisions shaped it, and what remains uncertain.
Three commitments
Research works when the whole team can see it, shape it, and pick it up.
Most AI tools are designed around a conversation between one person and one agent. Research follows a shared trajectory, where every participant brings different evidence, judgment, and goals. AI must therefore align with the whole team and the project’s history, so each contribution builds on what came before instead of pulling the work in several different directions.
Keep the work together
Nowadays, researchers constantly switch across Zotero for literature, Obsidian for notes, Overleaf for writing, VS Code or Vim for code, remote servers over SSH, notebooks for experiments, and coding agents that operate across some of them. Now each tool is adding its own AI, and every agent inherits that tool’s boundaries. Instead of one project context, you end up with several isolated context making the agent do stupid mistakes. Of course, we avoid this by repeatedly re-explaining what it might miss, the flow of thought completely lost and leaving the fine-grained context to rot and practically disappear in each of the windows.
Terminal agents and connector-based apps have already shown how much more capable AI becomes when it can work across several systems without constant switching. But connection is not the same as control. A remote connector can change after it is approved, and the data it returns can carry instructions into an agent’s context. A one-time permission grant is not a durable security boundary. The research project itself should become the integration layer, with its artifacts in one workspace and access that is visible, revocable, and scoped to the project.
Make research multiplayer
In many research projects, collaboration still amounts to circulating the latest PDF. A coauthor comments on the manuscript, a PI reviews the final plots, and their feedback arrives after most important decisions have already been made. AI now lets each person move faster inside their own slice of the work, which can leave the rest of the team even further behind.
Multiplayer research means sharing the chain that produced the paper, not only the paper itself. Coauthors should be able to follow the hypothesis, literature, limitations, experiments, failed attempts, and decisions as they develop. With that context, every collaborator and their AI can understand the project closely enough to bring their expertise at the moment it matters, not only when the manuscript is ready for submission.
Make progress easy to pick up
Once research becomes truly multiplayer, the next challenge is keeping its trajectory legible. A result alone cannot tell a collaborator why one approach was abandoned, which assumptions remain open, or what a person or agent changed since they last looked. Without that history, returning to the project means reconstructing it before anyone can move it forward.
Progress is not a clean line from hypothesis to paper. It includes failed experiments, rejected explanations, changes of direction, and the evidence that justified them. Keep those branches connected to the current work, together with who or which agent made each contribution and what was checked. A coauthor can then see where their expertise matters, a PI can evaluate the direction, and the team can build on what it has already learned instead of repeating it.
Research should leave understanding behind
These three commitments lead to one standard: research should preserve enough of the work for someone to understand it. That means keeping the evidence behind a result, the approaches that failed, the decision that changed direction, and the limits that remain.
AI can produce an answer faster than a researcher can absorb the reasoning behind it. That makes attention a research resource. Delegate what is already understood, but slow down where judgment is still being built. Completing a task and understanding it are not the same achievement.
A legible project gives that attention somewhere useful to go. A student can study the assumption they do not yet understand, a collaborator can challenge a limitation, and a PI can apply experience while the direction can still change.
The goal is not to document research for its own sake. It is to make clear where human attention matters and to leave enough understanding for the next person to continue the work.
A paper should be the clearest explanation of a result, but it should also open the project behind it. Connecting each claim to the relevant data, code, environment, decisions, limitations, and failed attempts gives collaborators and readers a path to inspect, reproduce, challenge, and extend the work. When some material cannot be shared, those limits should be explicit.
Kosmico
This is why we are building Kosmico
We are building Kosmico around a simple idea: the project, not the paper or an agent conversation, should be the shared object of research.
Keep the work together. Make research multiplayer. Preserve the trajectory. When those three commitments meet, every researcher and agent can understand what came before, see what remains open, and contribute to what happens next.
We do research to learn and create knowledge that can outlive the team that made it. The paper can remain the explanation. The project should become the record that others can inspect, challenge, reproduce, and build on.