What problem was I trying to solve?
Whether an end-to-end ML research pipeline — natural-language question, plan, design, code, execute, analyze, write — could run on its own with a human intervening at the points that actually matter.
What did I build?
A complete CrewAI implementation of that pipeline, a React streaming UI against it, five human-approval gates, and fourteen architecture decision records. A pre-registered benchmark protocol was written and committed. Ports to AutoGen and LangGraph were started.
What actually happened?
The pipeline ran end to end and produced papers and metric files from a natural-language topic. The benchmark protocol was never executed. The AutoGen and LangGraph ports never ran end to end. No cross-framework comparison was ever performed.
Earlier versions of this repository opened by calling MARS the first open-source system to implement and compare the same ML research pipeline across three frameworks. That claim is withdrawn. Every framework invocation in the completed implementation constructed a crew of one agent running one task, and the largest phase never imported the framework at all — so the independent variable the comparison depended on was not instantiated in code.
The repository was archived read-only in August 2026. The teardown that followed is Agent Harness Anatomy.
What did I learn?
Without a step that checks — by counting the code — whether a study’s independent variable is actually instantiated, it is possible to spend months on research that is not merely unfinished but structurally untestable.
What can be verified?
ARCHIVED.md records the archive date, final version, final commit hash and commit count.
The README carries a component-by-component table of what was complete and what was not.
The pre-registration is in the repository, unexecuted. The three prototype directories and
their comparison scaffolding are present and countable.
Code
- Repository (read-only)
ARCHIVED.md- Successor:
agent-harness-anatomy