Closed-loop learning in autonomous knowledge-worker agents

Abstract

We present an architecture for autonomous knowledge-worker agents that continuously improve task performance through closed-loop learning. The system records execution artifacts — tool outputs, intermediate decisions, error signals, and outcome assessments — and feeds them back into the planning phase as structured memory. Over successive task executions, the agent learns which strategies succeed in specific contexts, reducing redundant discovery steps and adapting to organization-specific tool configurations. We evaluate the approach on multi-step enterprise workflows spanning CRM updates, document extraction, and cross-application data reconciliation, demonstrating measurable improvements in completion rate and execution efficiency across repeated task classes.

Key contributions

• Memory architecture that converts execution artifacts into reusable planning context
• Strategy recording with cache-hit-level tracking for progressive skill acquisition
• Sparse-to-dense tool input evolution that minimizes redundant discovery on repeat tasks
• Evaluation on real enterprise workflows across CRM, document, and reconciliation domains