How iquestlab/modularrsi works
ModularRSI is a research system for improving agent harnesses from execution experience and evaluating whether improvements transfer. Its documented workflow samples benchmark-independent trajectories, contrasts successes and failures, evolves harness modules independently, filters candidates through validation gates, integrates cross-module changes, then freezes the resulting function library for downstream evaluation. The tree exposes the included harness modules, but not the evolution and evaluation orchestration; those stages are therefore represented as documented subsystems rather than inferred implementations.
Components 16
Learning workflow
- Trajectory sampling
- Contrastive analysis
- Module-wise evolution
Harness modules
- Agent loop
generations/merged_active/gen_0/modules/agent_loop/baseline.py - Observation
generations/merged_active/gen_0/modules/observation/baseline.py - Tools
generations/merged_active/gen_0/modules/tools/baseline.py - Context management
generations/merged_active/gen_0/modules/context_mgmt/baseline.py - Completion detection
Validation and integration
- Program checks
- Diff review
- Execution validation
- Cross-module integration
Evaluation
- Frozen function library
generations/merged_active/gen_0/modules - Downstream evaluation
- Transfer evaluation
Other components
- Researcher
Connections 17
- Researcher → Trajectory sampling: starts evolution
README.md - Trajectory sampling → Contrastive analysis: provides trajectories
README.md - Contrastive analysis → Module-wise evolution: guides changes
README.md - Module-wise evolution → Agent loop: evolves module
README.md - Module-wise evolution → Observation: evolves module
README.md - Module-wise evolution → Tools: evolves module
README.md - Module-wise evolution → Context management: evolves module
README.md - Module-wise evolution → Completion detection: evolves module
README.md - Module-wise evolution → Program checks: screens candidates
README.md - Module-wise evolution → Diff review: screens candidates
README.md - Module-wise evolution → Execution validation: screens candidates
README.md - Program checks → Cross-module integration: permits integration
README.md - Diff review → Cross-module integration: permits integration
README.md - Execution validation → Cross-module integration: permits integration
README.md - Cross-module integration → Frozen function library: produces library
README.md - Frozen function library → Downstream evaluation: evaluated on
README.md - Downstream evaluation → Transfer evaluation: measures transfer
README.md