"My Coding Agent Remembered Sessions, Not Work. That Was the Bug"
The article, published on DEV Community, details a common bug in the author's coding agent: it remembered conversation sessions but not the work context. The agent was built on a simple architecture of conversation → runtime session → messages, which works for demos but fails in real-world usage. For example, starting a task from a web UI, continuing on mobile, approving commands, and later checking progress led to the system guessing from recent messages. It knew there was a session but not which job it belonged to. The author highlights scenarios like 'continue the routing task,' 'use Claude Code to review what Codex just changed,' or 'retry that, but keep the same working directory'—none of which are about chat sessions but about work. A runtime session can crash, users can switch from web to Telegram or Feishu, and multiple agents can work on the same issue from different roles. Treating the runtime session as the main identity breaks in all these cases. The article underscores the difference between a chatbot and a working assistant.
Agents must track work, not just sessions, to handle real-world multi-channel workflows.