{
  "id": 18899931,
  "title": "A harness grown from losses",
  "url": "https://urgent.news/2026/11/08/a-harness-grown-from-losses",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-08T08:50:44.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/serhovskyi/a-harness-grown-from-losses-jg1"
  },
  "original_language": "en",
  "account": "The harness, a collection of layers that decide what the model sees, does, learns, and stops, has evolved through a series of commits since its inception on August 14, 2026. The harness, housed within the Pantensa repository, consists of context, permissions, feedback, orchestration, and tests of the harness itself. The first five hooks, designed to refuse five past losses, were implemented immediately, followed by a two-day delay before the addition of another five lessons transformed into rules. These rules were tested and implemented just four minutes later. The harness operates on the principle that the environment, rather than the prompt, dictates the agent's behavior. It consists of five layers, each with a corresponding file in the repository, and is designed to minimize the loss of information during use. When a session begins, the harness measures the model's position and measures the memory index at 6144 bytes, as this file is loaded into every session. Permissions are enforced by a guard that prevents irreversible losses, such as commit, merge, rebase, and branch switching. Feedback is provided through warnings for commands that may produce false or misleading results, while orchestration ensures that long-running tasks are managed efficiently, with only the summary lines and exit codes from subagent transcripts being used. The harness closes a link by retaining the ledger and git log, rather than the dialogue, to minimize the context that the model has access to. In one instance, a single test file generated 74 MB of stderr during a single run, but only five lines of this output were retained in the context.",
  "summary": "I almost never see the auto-compact warning in Claude Code. That is not a setting. I dislike what compaction does — the conversation is replaced by a lossy summary, and the nested instruction files do not survive it — so the workflow was built so that almost no session reaches half of its context window. The \"half\" is an observation, not a measurement: nothing I keep records how large a session's…",
  "key_points": [
    "Harness evolved through series of commits since August 14, 2026",
    "First five hooks refuse five past losses, added immediately",
    "Harness closes session by retaining ledger and git log"
  ],
  "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."
}