编程 Agent 增长遭遇人类评审瓶颈:Tessl 博客解析信任赤字

来源:Tessl:产品与工程博客 2026年10月20日 00:00 AIHOT 评分:69 精选
摘要:Tessl 博客引用 Jellyfish 2026 年 9 月 AI Engineering Trends 数据指出,中位数公司已有 22% 的合并 PR 由自主编码 Agent 撰写(90 分位达 53%),但仅约 0.6% 的 PR 由 Agent 评审合并(前 10% 公司为 9%)。
推荐理由:基于 Jellyfish 平台数据指出 agent 写码占比与人工评审瓶颈的落差,并归纳领先团队用平台工程加速建立信任的做法。
# 编程 Agent 增长遭遇人类评审瓶颈:Tessl 博客解析信任赤字 > 原标题:There Aren’t Enough Chaperones for All Your Coding Agents - 来源:Tessl:产品与工程博客 - 发布时间:2026-10-20T16:00:00.000Z - AIHOT:https://aihot.news/items/wz8yuzgpztob6gtdo8pxuip2j - 原文:https://tessl.io/blog/there-arent-enough-chaperones-for-all-your-coding-agents ## 摘要 Tessl 博客引用 Jellyfish 2026 年 9 月 AI Engineering Trends 数据指出,中位数公司已有 22% 的合并 PR 由自主编码 Agent 撰写(90 分位达 53%),但仅约 0.6% 的 PR 由 Agent 评审合并(前 10% 公司为 9%)。 ## 推荐理由 基于 Jellyfish 平台数据指出 agent 写码占比与人工评审瓶颈的落差,并归纳领先团队用平台工程加速建立信任的做法。 ## 正文 · 中文译文 想象一下,你领导着一个工程团队,而你对其中 20% 的开发者存在严重的信任问题。 这些开发者足够可靠,你会让他们写代码,但他们需要相当多的监督。他们不能独自担任他人工作的审查者,当然也不能批准将代码合并到重要系统中。就开发者而言,他们属于二等公民。 传统软件工程会将这种状况视为严重的结构性问题。高效的开发团队建立在信任高自主性个体的基础之上——这些人(配合良好的流程和保障措施)能够编写、审查和交付代码,而无需被微观管理。 然而,这正是当今典型工程团队所处的境况,只不过我们所说的这 20% 开发者是 AI 智能体。大多数团队要求每个智能体都有人陪同——而陪同人员根本不够用。 ## 智能体信任问题,两张图表 在 Jellyfish,我们追踪工程组织如何使用 AI 编码工具和智能体,以及这种使用如何与产出相关联,比如实际交付了什么代码、构建了哪些产品和功能,以及这些事情的成 本是多少。通过我们的平台,我们可以洞察约 1,300 家公司中的约 30 万名工程师和超过 1 亿个拉取请求,这是业内最大的软件工程工作流数据集。 在最近一期的月度 [AI 工程趋势报告](https://jellyfish.co/ai-engineering-trends/)中,有两张图表尤其凸显了工程团队在扩大自主编码智能体使用规模时面临的挑战。 第一张是自主智能体活动,我们将其定义为一 家公司已合并 PR 中由自主编码智能体编写的比例(即由编码智能体编写代码、提交和/或发起 PR)。中位数公司目前为 22% 的 PR,高于一个月前的 18%。处于第 90 百分位的公司为 53%。 _按公司百分位划分的自主智能体活动。来源:Jellyfish AI 工程趋势,2026 年 9 月。_ 对于任何在 2026 年使用过 AI 编码智能体的人来说,这些工具的惊人普及率应该不会太令人意外。执行框架和模型能力的进步使 AI 编码智能体能够接管很大一部分编码工作负载。但可以说,这只是“原始编码”部分——你仍然需要批准、合并和部署这些代码。 让我们看另一张图表:“智能体合并的 PR”,即由智能体审查并合并、未经人工审查的拉取请求。这种做法远不那么常见。中位数公司仅对约 0.6% 的 PR 实现了纯 AI 审查,而排名前 10% 的公司为 9% 且仍在攀升。 _按公司百分位划分的智能体合并 PR。来源:Jellyfish AI 工程趋势,2026 年 9 月。_ 因此,在典型公司中,智能体现在编写了约五分之一的代码,但几乎所有这些代码在流程的某个环节都需要人工介入。这种人工监督是建立信任和确保质量的合理方式,但它也构成了编码智能体能够提供多少额外产能的硬性约束。 大家都在谈论软件工厂,但一个每个单元在出厂前都必须经过人工检验的工厂并不是工厂——它只是一个拥有更多传送带的工作坊。 ## 审查瓶颈是一个 ROI 问题 评审瓶颈不仅让工程师们倍感压力,也让领导团队倍感压力,因为它是造成 ROI 问题的一大因素。 所谓“ROI 问题”,我指的是上游代码创建速度的提升远比下游影响更为显著。例如,在 2026 年全年,我们普遍测到拉取请求(PR)数量增长了约 2 倍,但已交付功能等产品产出仅增长了约 40-50%。 造成这一差距的原因之一是,由智能体编写的 PR 在早期阶段就失败的比例要高得多。当由人类编写 PR 时,约有 20% 的情况无法合并,原因各种各样且完全正常(例如实验性工作、方案错误、新需求、优先级变化)。而对于智能体,放弃率翻了一倍(40% 对 20%)。与此同时,那些确实合并了的智能体 PR 反而受到更多审查而非更少:过去一年里,每个已合并智能体 PR 所获得的人工评审数量增加了一倍多。 这些本身都并非不合理。智能体代码确实更难评审,评审者对其保持谨慎是正确的。问题在于,随着编码智能体能力持续增长,这些负担给整个系统带来的压力。评审负担随着智能体产出的增加而扩大,而人类注意力始终是固定的约束,并且我上面展示的自主性曲线每个月都在复利式增长。 如果今天处于中位水平的公司达到今天前 10% 公司的水平(即自主智能体编写超过一半的代码),代码评审就有可能变成人类的全部工作,从而挤占经验丰富的开发者本应从事的其他所有重要事项的时间,比如与产品和设计协作、塑造软件架构,以及将他们的工程判断力用于那些没有明确解决方案的问题。 ## 如何更快地建立信任 幸运的是,一些公司正在解决这个问题。那些让自主智能体合并 9% PR 的顶尖公司,并不是在盲目地更加信任它们的智能体。相反,它们投入精力来加速信任的建立速度。 这通常涉及大量平台工作,例如将低风险工作路由给自主智能体的系统、用于安全可靠开发与部署的沙箱和预发布环境,当然还有大量的自动化测试。不过,也许最重要的是,它们很早就开始并大力推进;信任只能随时间建立,你等待得越久,就要越久地忍受不断堆积的评审积压。 **_对每个智能体全程陪护是一种合理的起步方式,但无法规模化。_** 如果你有兴趣进一步了解自主智能体如何演进(或者更准确地说,如何颠覆)SDLC,欢迎于 11 月 2 日至 4 日在纽约市参加 [aiDevCon](https://tessl.io/devcon)。我将在我的演讲_《Botsitting Doesn’t Scale: How to Navigate the Autonomy-Trust Gridlock in Real Engineering Orgs》_中谈论这一点以及更多内容,并分享我们从分析 Jellyfish 数据中学到的内容:最优秀的工程团队如何应对 2026 年最大的工程工作流挑战。 ## 正文 · 原文 Imagine leading an engineering team where you have some major trust issues with 20% of your developers. These developers are reliable enough that you let them write code, but they require considerable supervision. They can’t be the sole reviewer on other people’s work and they certainly can’t grant approval to merge code to important systems. As developers go, they’re second-class citizens. Traditional software engineering would view this state of affairs as a serious structural problem. Effective development teams are built on a foundation of trust in high-agency individuals who (along with good processes and safeguards) can write, review, and ship code without needing to be micromanaged. However, that’s the situation the typical engineering team finds itself in today, except the 20% of developers we’re talking about are AI agents. Most teams require each agent to be chaperoned – and there aren’t enough chaperones to go around. ## The agent trust problem, in two charts At Jellyfish, we track how engineering organizations are using AI coding tools and agents, and how that usage connects to outputs like what code actually ships, what products and features are being built, and how much these things cost. Through our platform we have visibility into about 300K engineers across 1,300 companies and more than 100 million pull requests, the largest body of software engineering workflow data in the industry. In a recent edition of our monthly [AI Engineering Trends Report](https://jellyfish.co/ai-engineering-trends/), two charts in particular underscore the challenges that engineering teams face today as they scale their use of autonomous coding agents. The first is autonomous agent activity, which we define as the share of a company’s merged PRs that were authored by autonomous coding agents (i.e. with coding agents writing the code, making the commits, and/or opening the PR). The median company is now at 22% of PRs, up from 18% just a month earlier. Companies at the 90th percentile are at 53%. _Autonomous agent activity by company percentile. Source: Jellyfish AI Engineering Trends, September 2026._ To anyone who’s been working with AI coding agents in 2026, the incredible adoption of these tools shouldn’t be wildly surprising. Advances in harness and model capability have allowed AI coding agents to take over a large proportion of coding workloads. But that’s just the “raw coding” part, so to speak – you still need to approve, merge, and deploy this code. Let’s look at another chart: “agentically merged PRs”, i.e. pull requests that were reviewed and merged by agents without human review. This practice is far less common. The median company has achieved AI-only review for about 0.6% of its PRs, while the top 10% of companies are at 9% and climbing. _Agentically merged PRs by company percentile. Source: Jellyfish AI Engineering Trends, September 2026._ So at a typical company, agents are now authoring about a fifth of the code, but almost all of it requires a human at some point in the process. This human oversight is a reasonable way to build trust and ensure quality, but it also represents a hard constraint on how much additional capacity coding agents can provide. Everyone is talking about software factories, but a factory where every unit has to pass a human inspector before it ships isn’t a factory – it’s just a workshop with more conveyor belts. ## The review bottleneck is an ROI problem The review bottleneck isn’t just stressing out the engineers, it’s also stressing out the leadership team, because it’s a big contributor to the ROI problem. By “ROI problem” I mean that the upstream acceleration of code creation is much more dramatic than the downstream impact. For example, over the course of 2026 we’ve generally measured about a 2x increase in pull requests, but only about a 40-50% increase in product outputs like features shipped. One reason for this gap is that agent-authored PRs die on the vine at a much higher rate. When a human authors a PR, it fails to merge about 20% of the time, for all sorts of totally normal reasons (e.g. experimental work, wrong solution, new requirements, changing priorities). For agents though, the abandonment rate is double (40% vs 20%). Meanwhile, the agent PRs that do merge are getting more scrutiny rather than less: over the past year, the number of human reviews per merged agent PR has more than doubled. None of that is unreasonable on its own. Agent code is genuinely harder to review, and reviewers are right to be careful with it. The problem is the strain these loads put on the system as coding agent capability continues to grow. Review load scales with increasing agent output while human attention remains a fixed constraint, and the autonomy curve I showed above is compounding every month. If today’s median company gets to where the top 10% are today (with autonomous agents authoring more than half of all code), code review threatens to become the entire job of the human, stealing time from all the other important things that experienced developers need to be doing like collaborating with product and design, shaping software architectures, and bringing their engineering judgment to problems that don’t have clear solutions. ## How to build trust faster Fortunately, some companies are figuring this out. The top companies that have autonomous agents merging 9% of PRs aren’t just blindly trusting their agents more. Instead, they’re putting in the work to accelerate how fast trust can be built. Often this entails a lot of platform work, such as systems for routing low-risk work to autonomous agents, sandboxes and staging environments for safe and secure development and deployment, and of course lots and lots of automated testing. Perhaps most importantly though, they’ve started early and are pushing hard; trust can only be built over time, and the longer you wait, the longer you’ll have to live with a mounting review backlog. **_Chaperoning every agent is a sensible way to start, but an impossible way to scale._** If you’re interested in learning more about how autonomous agents are evolving (or more accurately, disrupting) the SDLC, come join me at [aiDevCon](https://tessl.io/devcon) in New York City on November 2–4. I’ll be talking about this and much more in my session _Botsitting Doesn’t Scale: How to Navigate the Autonomy-Trust Gridlock in Real Engineering Orgs_, and sharing what we’ve learned from analyzing Jellyfish data about how the best engineering teams are tackling the biggest engineering workflow challenges of 2026.