XNews

AI Developments and Insights

AI Developments and Insights

Ben Clavié advocated for specialized Knowledge Agents over coding agents, emphasizing the importance of retrieval, task orchestration, and specialization in enterprise applications. Separately, a new open-source plugin 'claude-reflect' for Claude Code addresses agent memory loss by capturing corrections and adding them to configuration files through a human-verified queue.

Key insights

Knowledge Agents should prioritize retrieval and orchestration over coding capabilities

Ben Clavié argued at the AI Engineer World's Fair that agents for knowledge work should not be built like coding agents. Instead, they should focus on strong retrieval capabilities, effective task orchestration, and specialization for specific domains. The talk emphasized that the design of knowledge systems requires different architectural choices than those optimized for software development.

claude-reflect plugin solves agent memory loss by capturing corrections

The claude-reflect plugin for Claude Code automatically captures corrections when users correct the agent, praise its output, or say 'remember:'. The /reflect command presents a queue of these items for human review, and confirmed items are written to CLAUDE.md or other configuration files, ensuring all future sessions inherit the correction. It supports multiple config files (CLAUDE.md, subdirectory ones, AGENTS.md) and works with Codex and Cursor.

Plugin includes skill generation and multilingual support

Version 2 of claude-reflect adds /reflect-skills, which reviews past two weeks of conversations to identify repeatedly used phrases and suggests creating a command. The plugin uses AI semantic filtering to understand corrections even when given in Chinese, and displays confidence levels for each capture. All additions require manual approval, preventing unwanted configuration changes.

Source posts