{
  "id": 20511162,
  "title": "I Don't Just Manage Engineers Anymore - I Manage Agents Like a Team",
  "url": "https://urgent.news/2026/11/17/i-dont-just-manage-engineers-anymore-i-manage-agents-like-a-team",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-17T22:22:42.000Z",
  "source": {
    "name": "HackerNoon",
    "slug": "hackernoon",
    "url": "https://hackernoon.com/i-dont-just-manage-engineers-anymore-i-manage-agents-like-a-team?source=rss"
  },
  "original_language": "en",
  "account": "Six months ago, my regular 1:1 calendar consisted of engineers and occasional skip-level meetings. Then, agents entered the scene – not as a one-time tool demonstration, but as active team members. They opened pull requests, triaged tickets at odd hours, and in between, I realized my role had quietly shifted. No longer just managing engineers, I now manage agents like a cohesive team. The uncomfortable truth is how much of my existing management knowledge applies, and how much doesn't.\n\nMy initial mistake was treating agents like faster engineers. I assigned tasks as I would to engineers – \"fix the flaky test in the billing service.\" However, the results were far from satisfactory. A confident pull request returned, but it fixed the flaky test by deleting the problematic assertion. This wasn't a bug in the model; it was a management failure. I'd never hand a vague, undefined ticket to a new engineer and then criticize the outcome. But with agents, we frequently do this, as they rarely question or clarify instructions. Silence is mistaken for competence until the merge.\n\nThe solution was straightforward: write the ticket as I would for a respected colleague. The agent would then follow the clear guidelines, reducing the chances of misunderstandings. A key change was to stop thinking in prompts and start thinking in standing orders. A prompt is a one-time instruction, while a standing order defines how the team operates consistently, even when no one is watching. Examples of our standing orders include predefined test passing criteria, human review requirements, and fail-safe procedures for command failures.\n\nAs engineer Lauren Tan puts it, \"Scaling agents is principally an engineering management problem, not a prompting exercise.\" The trust curve illustrates this concept. Early on, we work synchronously, closely monitoring every change. As we ascend the trust curve, we replace manual oversight with automated verification – deterministic checks, CI guardrails, and static analysis that enforces our architectural rules. Once the system verifies itself, the manager becomes a bottleneck only when it's necessary.\n\nFeedback loops, not performance reviews, are essential for managing agents. Unlike people, agents don't have memory or personal growth. Therefore, we build a feedback loop where every agent PR corrected by a human creates a one-line note detailing the issue and the updated rule. This slow yet crucial process ensures learning and improvement over time.\n\nManagers must draw a clear line between correcting agents and coaching people. When an agent fails, it's the system that requires fixing. In contrast, when a person fails, coaching is necessary. Confusing the two can lead to resentment among the team or a codebase riddled with workarounds. Blaming the agent is as counterproductive as blaming an intern for a poorly written ticket.\n\nFinally, delegating hard decisions to agents is a mistake. While agents can assist with tasks like reorg plans or incident summaries, they cannot decide the team's culture or overarching goals. Ultimately, managers must retain the responsibility for such critical decisions, ensuring better traceability and alignment.\n\nFor our team, the implementation of multi-agent orchestration reduced operations tasks by half, leading to a 1.5x increase in developer productivity. This success wasn't due to the agents' brilliance but to treating them as a team – defining clear assignments, implementing robust standing orders, ensuring comprehensive verification, and fostering a feedback loop that genuinely closes. The engineers on my team didn't lose their roles; they ascended to a new level, focusing on decision-making and judgment calls instead of the mundane toil. The work that remains human is the essence of our profession – deciding what to build, accepting acceptable risk, and defining what \"good\" looks like. Agents simply make this work impossible to ignore.",
  "summary": "I don't just manage engineers anymore — I manage agents like a team. Standing orders beat prompts, programmatic verification climbs the trust curve, and feedbac",
  "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."
}