{
  "id": 14245825,
  "title": "An agent can suggest a lesson. Should it become a rule?",
  "url": "https://urgent.news/2026/10/13/an-agent-can-suggest-a-lesson-should-it-become-a-rule",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-13T16:53:40.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/mottych/an-agent-can-suggest-a-lesson-should-it-become-a-rule-2nl7"
  },
  "original_language": "en",
  "account": "When using an agentic workflow for coding, it is common to encounter situations where an AI coding agent may become stuck due to minor, seemingly irrelevant issues. For instance, in a recent scenario, an agent failed because it lacked Jam credentials. Although a human would not dwell on this issue, the agent would until instructed otherwise. These sorts of corrections can serve as valuable lessons for future runs; however, turning each correction into a permanent instruction can lead to an accumulation of rules with no one monitoring their relevance.\n\nTo address this issue, a proposed solution involves allowing agents to submit learning proposals. These proposals are reviewed by a librarian agent, which then generalizes them into readable candidate rules, taking into account the original prompt's instructions. The librarian agent then presents these rules and their corresponding agent instructions to an operator for approval, editing, or rejection. This process ensures that the proposed rules are useful, specific, and applicable to the appropriate agents and stages.\n\nBy employing this system, the learning process becomes more efficient and targeted. It prevents an unnecessary accumulation of rules while still providing agents with useful guidance. Agents report when they utilize a rule, and the librarian uses this data to identify rules that are not frequently utilized, potentially leading to their retirement. Furthermore, when a proposal is rejected, the reason for rejection is recorded and used by the librarian for future proposals.\n\nUltimately, this self-learning system helps maintain instructions without overwhelming agents with excessive guidance. It strikes a balance between being too specific and too broad, ensuring that only relevant rules are passed on to future runs. However, the question remains: how much instruction should persist from one task to the next? If you utilize coding agents, what strategies do you employ to handle this issue today? Do you continuously add instructions, or do you also review and remove them?",
  "summary": "If you are running an agentic workflow for coding, I am pretty sure you already know that - sometimes an AI coding agent can get stuck for a stupid, boring reason. For example, in a recent run of my own workflow, an agent failed because it was missing Jam credentials. I use Jam to report bugs and provide a link and the MCP to the triage agent to help it understand the bug, but this run failed…",
  "key_points": [
    "Agents can suggest lessons as learning proposals.",
    "Librarian agent reviews and generalizes proposals into rules.",
    "Approved rules presented to operator for approval/edit/rejection."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}