The editor should learn the project, not just open it
Draft seed for the launch thesis post. Outline, not copy. Narrate judgment, not momentum (
docs/brief.md§devlog).
The bet
The unit of work moved from files to sessions. Agents — not hand-editing — are how work gets done now, and the session (prompt → decisions → diffs → result) is the real artifact. Every incumbent editor is still organized around the file.
Why now
- What changed: directing agents replaced editing files for a growing class of developers.
- What nobody’s built: a workspace organized around the session, with a layer that learns the project and compounds across sessions.
- The contrarian edge: persistence ≠ memory; hosting ≠ learning.
The proof
The one differentiator, working: the editor that learns the project, so each session starts smarter than the last. (Link the real loop, not a mockup.)
What it is / isn’t
Hosts the real agent CLIs in real terminals; surrounds them with orchestration, review, and learning. Not an IDE — no LSP, autocomplete, or refactoring.
Who’s building this
Pedigree → thesis as causation. A real way to reach + follow.
Routes to: repo · site · follow · early-access.