A new recurrent harness framework, ReHarness, is proposed for agentic continual learning. It combines forward and backward information flow to improve online adaptation to distribution shifts. ReHarness outperforms existing methods across three benchmarks and three language model backbones.