JIG
A modular runtime for agent loops and pipelines
A modular runtime for agent loops and orchestrator-controlled pipelines, with tracing, grading, feedback, and interchangeable adapters.
JIG provides the runtime for agent loops and explicit pipelines. Models, tools, memory, grading, feedback, and tracing connect through small interfaces so the surrounding system can control how work is executed and measured.
Two execution models
In an LLM-directed agent loop, the model chooses the next tool call and JIG runs the completion/tool cycle. In an orchestrator-controlled pipeline, the surrounding system defines the sequence, with tracing, grading, and feedback around each step. A model can be a transform within a step without choosing the workflow.
Provider, memory, grading, feedback, and tracing adapters support specialization through common interfaces. Integration coverage differs: SQLite memory and feedback form the verified baseline; the repository’s Honcho and Zep adapters are not covered by integration tests.
Runtime and surrounding systems
Scout uses a typed JIG agent to evaluate and draft. The coding experiments in Assay include pinned slices of JIG and Scout as study subjects. These are distinct relationships: JIG supplies runtime mechanisms, while the surrounding applications define tasks, experiments, and oversight.
Installing JIG does not implement the complete Progressive Autonomy Architecture governance lifecycle.
What to inspect
The README documents the interfaces and execution models with runnable examples. The Jev client guide covers typed feature-scoring questions and answers.