AN AGENT. A SHELL. A LOOP.
Born from my research workflow: the agent organizes intent, writes the code, and verifies the work.
As the experiments continued, the agent needed a persistent working environment.
So the core entered a shell—a home for knowledge, delivery topology, and workflows.
The boundary between shell and world begins to blur: through the shell, the core modifies the world; by modifying the shell, it modifies itself.
Understands intent, calls tools, runs experiments, and modifies code.
Loop harness and visualization, so the core's behavior is structured and observable.
ENTRY / THE SHELL
Users need a way to browse experiments.
Two machines, three repos, many branches: a structured CI and branching system
Research develops across stages into a coherent set of conclusions. This view presents the process and key results.
As the shell evolved rapidly from scratch, it accumulated loops, hard gates, constraints, and archives. A dashboard brings them together and makes ongoing shell modifications observable.
ORIGIN / FROM THE LAB
During my research, I noticed that the agent documented experiment procedures in specs, plans, and configuration files, and archived the results alongside them. Each experiment now had a structure I could inspect and run again.
The work spanned a local machine and a server, alongside research and knowledge repositories. I gave those operations to a dedicated agent: create the run, collect the result.
As experiments accumulated, I built a frontend to view configurations, results, and logs, and paired it with the agent. This created a distinction between shell and core: I interact with the core through the shell, using it to define gates and shape behavior.
Then the shell became a task for the core. The core modifies the shell in a sandbox and launches a separate agent to verify the changes. The changes are merged only after external review. With a sense of touch—the ability to sense user interactions—the core began interpreting those actions and forming intentions to modify the shell.
LOOPS / MULTIPLE LOOPS
Turn user intent into an experiment plan, execute, collect the result
Understand user needs through interaction signals and conversation, modify the shell in a sandbox, then submit the changes for external approval and merging.
Evaluate cost, safety, and effectiveness. Capture reusable test cases, run them, and review their results with one-click actions.
AGENT FACTORY / WHERE THIS GOES
Swap in another project's knowledge, a set of task SOPs, and verification rules. The core works in a new context, keeping the structure of observation, tests, review, and delivery.