返回资讯中心
外部精选
软件工程
#开发者实时流

The Death of the Backlog

You can learn a lot about a company by looking at how it decides what to build and what to punt. Embedded in that process are its hierarchies, priorities, and lines of influence. What makes it to the list reflects a com…

Hacker News 最新avyfain2026/11/17 · 今天17 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

You can learn a lot about a company by looking at how it decides what to build and what to punt. Embedded in that process are its hierarchies, priorities, and lines of influence. What makes it to the list reflects a company’s culture. I've spent much of my career building systems to manage this process, but the backlog as we know it is dying.

I started my career at Apple, where planning happens on a scale that's hard to appreciate from the outside. The whole company runs on a tight cadence of tentpole features, decided well in advance, and everything else falls into place around them: hiring, promotions, manufacturing supply chains, data center buildouts, and so on. Apple builds its own task tracker, Radar, which has bugs and feature requests going back to the 80s. If something doesn't exist in Radar, it might as well not exist. Paradoxically, it's both the central nervous system for much of the company and a black hole.

Moving to a startup was a shock. At Vouch, they were introducing Jira when I joined. Planning cycles were shorter, but the backlog served the same purpose: Operations, Underwriting, Marketing, and Sales funneled requests into spreadsheets for Product to prioritize with Engineering. Someone had to figure out what depended on what, and decide which would move the metrics each cared about. We'd rank everything on the classic two-by-two of impact and effort, negotiate across teams, and eventually hand the work to engineers to break into smaller pieces.

At AngelList, larger and storied, I found almost the opposite. There wasn't a formal bug-tracking system for anything outside of on-call work. Jira and Linear were both present, but only some leads organized their work in siloed Notion pages. Engineers mostly picked up work as it came in, from Slack. I pushed to introduce capacity planning, leaning hard on Linear to cut requests into projects and tickets, recording organizational commitments. At first just my team tried variations of kanban and agile, which led to arguments with engineers used to the old way. Backlogs are where tickets go to die, they'd say. And that's a good thing, I'd respond. It's better to let a ticket die than waste time on something that doesn't matter. I'd written about this years earlier: a big part of an engineering manager's job is to find and remove bottlenecks, filtering requests and deciding what deserves attention so teams can get more meaningful work done with the same resources.

The industrial engineer in me spoke up: if your team can do twenty things in a week but fifty requests come in, what happens to the other thirty? When your throughput is lower than the rate of arrival, you need to write down what you're not doing and be deliberate about what you pick up next. A backlog is inventory, and keeping inventories is costly. Writing tickets, organizing them into groups, constantly revisiting priorities, and cleaning up requests that have become irrelevant after sitting around (sometimes for years) is a full time job. The overhead made sense when engineering time was expensive and getting something wrong could mean weeks of wasted effort.

Historically, we spent a lot of time deciding whether something was worth building because building it was expensive. For a growing class of software problems, the cost of asking an agent to take a first pass is comparable to the cost of writing a decent ticket. It's not free, but certainly cheaper than seeing the same ticket come up every sprint planning session, spending time gathering details on how the original request might have changed, or tracking down the current owner of that problem when someone leaves the team, only to eventually decide it wasn't worth doing. Increasingly, you just try it. Maybe the agent produces something terrible and you chuck it. But maybe it takes twenty minutes and turns out to be easier, or more valuable, than anyone expected.

And yet not every request deserves to be built. If the goal is to go West and someone asks to go North, you should still say no. Cheap distractions are still distractions. But if we agree an idea might move the right metric, why not? The question changed from “can we lift it to this sprint?” to “is there any reason not to do it now?”

I've spent the past few months building, working by myself with a team of agents. It's changed how I think about software development. Working alone eliminates most of the coordination costs that come with having a team. At first, I kept a to-do list, and tried to have agents consume from it, but quickly realized I didn’t need it. The rate at which we build things has become higher than the rate at which we can come up with ideas. That doesn't mean throughput is infinite, though. The bottleneck just moved elsewhere, as it does, and I’m sure as soon as I bring on the first few hires it’ll move again. If engineering output increases tenfold, review, integration, and product judgment become relatively scarcer.

Of course, working alone is a special case. A team of agents working for one person doesn't have the competing priorities or organizational politics of a team of humans. But working with agents has actually forced me to adopt processes I'd never have bothered with on solo projects: PRs, CI, automated testing, and so on. You can eliminate the humans from an engineering team while increasing the need to coordinate the work.

There's also a difference between just trying something and shipping it. A company that builds everything requested will accumulate a remarkable amount of garbage, and the coordination problem might actually get worse when people start to depend on that garbage. And maintaining that software will eventually be someone’s job.

Planning is not going away. Big projects still need coordination, and someone still has to decide where the product is supposed to be going and who’s doing the building. As that constraint loosens, more people throughout an organization will be able to test their own ideas without first convincing someone to allocate engineering capacity. A deep company brain might be able to accumulate the right context for the agents querying it to reason about the way their tasks ladder up into the bigger goals, reducing coordination cost, but it might also send those agents astray. I expect AI native companies will spend a lot less time maintaining inventories of things we might someday build, and a lot more time figuring out what to try now cheaply and which of the things we've already tried are worth shipping.

Thanks to Hannah Doherty and Max Faingezicht for reading drafts of this post.

Photo: L'horloge d'Orsay, by me. Previously posted in Le Trip to France, 2026.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →