Research Radar · Experimental implementation
Agent Harness
Compositional execution-state management for long-horizon agentic code generation.
What it investigates
Agent Harness examines how an agentic coding loop can retain dependencies and selectively rerun work after feedback.
Implementation record
Its ExecutionStateGraph represents code steps and data dependencies as a directed graph. Documented modules use AST analysis to infer reads and writes, execute steps in namespaces, record outputs, mark work stale, interpret failures, and generate targeted patches before replaying a subgraph.
Public record
The linked research item is the public source recorded for this Research Radar build. It provides context for the implementation; this page does not claim to reproduce its results. The repository and its documentation are the current public materials for inspecting the build itself.
Scope and limitations
The implementation is a reference for these execution-state mechanics. It does not demonstrate reliable long-horizon code generation in arbitrary repositories or independently validate patch quality.