Research Notes
Three reading logs. One paper per entry, written up as I go.
RSI · Recursive Self-Improvement
Can a system improve the machinery that produces its own intelligence?
Daily notes on agents that improve their own scaffolding — self-referential optimisation loops, evolved harnesses, and the empirical question of whether any of it compounds. Roadmap follows Lilian Weng's harness-engineering survey.
CLI · Agent-Native Interfaces
Should agents act through the pixels of a human GUI, or through a constructed command surface?
How agents should operate software. Controlled GUI-vs-CLI comparisons, skill compilation, MCP tool layers, and the recurring finding that the interface can outweigh the model.
RUA · Robot-Use Agents
Is a coding agent already a visuomotor policy, if you hand it the robot's own interface?
Agents that operate a robot the way they operate a computer. Code as policies, agents wrapped around frozen VLAs, ROS 2 as the native action contract, and the benchmarks that show how little of the reported progress is actually zero-shot.