{
  "id": 12946805,
  "title": "Part 6: An operating system for coding agents: the disciplined build",
  "url": "https://urgent.news/2026/10/07/part-6-an-operating-system-for-coding-agents-the-disciplined-build",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-07T20:09:22.000Z",
  "source": {
    "name": "Stack Overflow Blog",
    "slug": "stack-overflow-blog",
    "url": "https://stackoverflow.blog/2026/10/07/part-6-an-operating-system-for-coding-agents-the-disciplined-build/"
  },
  "original_language": "en",
  "account": "The first coding agent session is concise, reading code, making changes, and writing tests. After 30 attempts, the agent fails, re-discovering files, contradicting decisions, guessing test commands, and stopping test writing. The fix is not a better model; instead, build an operating system around the model: a constitution, autonomy boundary, tiered context, durable memory, review panel, and post-edit hook. The model's influence on whether the agent produces working software is minor, but the harness (system prompt, tools, context policies, hooks, sandboxes, sub-agents, observability) is crucial. Changing the harness, like LangChain's system prompt, tools, and middleware, can dramatically improve performance, as seen in the Terminal-Bench 2.0 benchmark where LangChain boosted the coding agent's performance from 52.8% to 66.5%. Most agent failures stem from configuration issues, not the model itself. A lean harness file outlines a workflow loop, autonomy boundary, quality bars, and adapter discipline (no vendor SDK in domain code). The essential decision is to allow the agent free rein on reversible changes while pausing on irreversible ones like pushing to remote and deploying. A giant context file can overwhelm the system; thus, tier the context, fixing stale context within the same change. Use focused sub-agents to review parallel changes, employing high-confidence gate reviews. Coordination between agents is achieved through shared files encoding the state of the work, rather than conversation, making it scalable and auditable. Ownership maps, claim/release rules, progress logs, and decision logs help manage agent interactions and maintain a record of decisions. This harness allows agents to remain productive for months, even when running multiple agents on the same repository.",
  "summary": "Coding agents are great for an afternoon and mediocre for a quarter. Here’s the small set of files and rules — with the actual configs — that keeps quality from decaying across months and thousands of edits, whether you run one agent or a fleet of them on the same repo.",
  "key_points": [],
  "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."
}