How-to

从全自动到人机协作:工程师反思用 coding agent 的常见误区

Why we picked it

Saravia 分享了他用 coding agent 的踩坑经验:别想着全自动,要去设计带 verifier 和 evals 的"loops"。和市面上鼓吹全自动 agent 的思路不同,他主张人更深地参与。天天写代码或做研究的开发者都能直接照着改进自己的工作流。

Elvis Saravia(omarsar0)发帖反思自己早期用 coding agent 的一个错误:试图一次性把所有环节自动化,结果依赖一堆主观、粗糙的 skills,效果不佳。

Body

他现在把大部分时间花在帮助 coding agent 写更好的代码、探索研究想法上,人反而更深度参与。他把有效的做法沉淀为"loops"(循环):循环里既包含自动化,也包含可持续打磨的明确步骤,通过 verifiers、确定性步骤、上下文格式化、记忆系统和 evals 来控制质量与稳健性。他表示这种方式明显降低了错误率。

他认为工作重心已经从"用 LLM 帮我们写代码"转变为"帮助 LLM 写得更好更快",coding agent 本质上是一个需要精心调校的传统系统。他还指出 coding agent 大约在一年前才真正起步,agent swarms 等核心原语尚未成熟,建议深耕自己的专业领域以应对明年变化。

Original · elvis

Software engineering and research have changed significantly for me. A big mistake I made early on was trying to automate everything. Instead of trying to automate everything at once with a bunch of silly and subjective skills, I now spend most of my time helping my coding and research agents write better code and explore novel research ideas. That's right! I am more involved. I think this approach works best today, given the current state of coding agents. Not only for coding but for researching as well. This process has resulted in good patterns codified as loops. Loops involve automations, but they also include clear, intentional steps you can keep refining to remove jaggedness. I continually strengthen my loops through verifiers, deterministic steps, context formatting and awareness, memory systems, and evals that steer quality and robustness. So I agree with other great engineering leaders that our job has fundamentally changed. It's not about using LLMs to help us write code. It's helping the LLM write code better and faster. That's where those loops come into play. This has significantly reduced error rates and allowed me to expand and scale with coding agents in unbelievable ways. Something I naively expected from my initial, silly, outdated automations. It feels contradictory at first, since we treat coding agents as solve-it-all intelligence, but it's a system like other traditional systems that requires careful attention to unlock its benefits. Like my other peers, I believe we are in the early innings. There are more core primitives, interfaces, and experiences to build. For example, we still haven't fully figured out agent swarms. Coding agents only really started to take off around this time last year. It's unbelievable how much has changed since then. To be better prepared for what's coming next year, I continue to advocate for gaining deeper knowledge and developing mastery and expertise in your domain. That will continue to matter.📊 19 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "How-to" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include coding agent, prompt engineering, agents, workflows.

Source attribution: curated from elvis (@omarsar0). Cross-source corroboration: 2 independent sources have covered this development (including AI Engineer), confirming industry-wide signal strength.