easy-coding-agents

About

easy-coding-agents is built around a failure mode common to autonomous loops: the model loses track of what it already established, repeats itself, or declares done without evidence. The engine's guards detect repeated tool calls, over-exploration of simple tasks, and empty responses, while a final-answer quality gate blocks DONE claims that lack verification evidence. The evidence-gated memory layer persists refs, tool logs, and task state across context wipes, and the repo publishes reproducible benchmark snapshots showing memory-subsystem recall after context wipe against summary, FTS, and vector baselines. It is a single-developer research project for people studying agent memory, not a production harness.