If you are a researcher who wants to study how people work with AIs to build software, what is your strategy for experimental design? It occurs to me that Bram could serve as a useful research instrument in various ways.
Using a guided workflow
Bram enforces a disciplined way of defining and executing subtasks (the “Worklist”). The most consistent feedback about Bram, so far, has been some version of: “I feel like I’m in control.” One possible A/B experiment: for a given assignment, group A defines and executes subtasks in a free-form way and group B uses the Bram workflow. How do participants feel about their control over the process?
Working in a repo
Bram manages work in a git repository and structures the work so that it flows into the repo in a clean and orderly way. Another A/B experiment: group A clones an unmanaged repo and performs git ceremonies in free-form dialog with agents, group B clones the same repo and uses Bram to direct git ceremonies. What do the resulting repos say about the quality and consistency of outcomes? What do subjects report about the cognitive effort required?
Searchable history
Bram captures four sources of information in its embedded SQLite database: git commits, GitHub issues, agent sessions, and Worklist history. When an experiment happens in a Bram-managed repo, fulltext search is available across the mixed corpus and can be used to ask questions whose answers draw on any or all of the sources.
Second opinions
Recruiting agents to review one another’s work is an emerging best practice. Bram supports Claude Code, Codex, OpenCode, and Antigravity and, in the context of a Worklist item, it is a one-click operation to ask one agent to review a plan or an implementation written by another. An A/B experiment: group A’s brief includes asking agents to provide second opinions, group B’s doesn’t. What do groups A and B say about their confidence in the results?
Reproducible scenarios
Bram itself is an app that mixes graphical and conversational user interfaces in complex ways. Testing of a new feature happens in a scratch repo that Bram spins up, populates with Worklist items designed to exercise the new feature, and monitors by watching an extensive and semantically-rich log. An A/B experiment: group A launches Bram in a repo with a Worklist that is structured in more granular and well-defined way than group B’s repo. What do the recorded results say about the quality and consistency of outcomes? What do subjects report about their feeling of control and understanding?
I am not a researcher!
So please take all this with a grain of salt. What I do know, for sure, is that we have a lot to learn about how people do — or could — work effectively in symbiosis with our new agentic partners. If you think Bram can help with the necessary research, let’s talk. I would love to help.
