The interesting curve - Human versus AI fleets一条耐人寻味的曲线：人，与 AI 舰队
2026-06-28
The curve that surprised me
I added a new line to the chart on homie-lab.com this week, almost as an afterthought. The home page only ever counted the async fleet's merged PRs, so the busiest days — the ones where the main agent and I shipped a big feature straight to master — showed up empty. That bugged me. So now every commit on master gets sorted into one of two buckets: part of an agent's PR (the async fleet), or a direct push (me + the main agent I actually sit with). Same unit, commits, side by side.
Then I looked at the two lines together, and the shape told a better story than I expected.
The comparison
Thirteen days in (since the fleet went live around the 14th), here's where it stands:
Commits landed on master, 6/14 → 6/28
Me + main agent ███████████████████████████████ 223
Async fleet █████████ 69
~3 to 1
Throughput, same window
async-fleet PRs merged ......... 69
human + main-agent commits .... 223
async-agent commits ............ 69
tasks triaged into backlog .... 149
The fleet's line is a low, steady drip — a few commits most days, never zero, rarely a spike. My line is spiky: long flat stretches, then a cliff-face of commits on the days a real feature lands. The 21st alone was a wall. (The live version animates day by day if you want to watch it move.)
That spread — roughly three to one — is the whole point, and it's the opposite of the thing people worry about.
What the shape actually means
The async agents are not replacing me. You can see it in the curve: they own the flat, boring baseline. Small fixes, stale docs, an untested screen, a flaky empty-state — the maintenance and the small-to-mid tasks that used to quietly rot when one tired person ran a real product alone. That's the steady drip.
The cliffs are mine. The main agent and I still take the hardest, most ambiguous work end to end — the architecture, the "what should Homie even be" calls, the things that need judgment. The fleet doesn't climb those cliffs. It clears the runway so I can.
And here's the part I didn't predict: handing off the baseline didn't shrink my output — it unleashed it. My 223 isn't 223 despite the fleet; it's 223 because of the fleet. When I'm not bleeding hours on chores, the scarce focus time goes entirely to building, and I build more, and faster, and weirder. The async line going up made my line go up too.
What we actually built
Two weeks, split by who did what:
Me + the main agent (the cliffs): - Homie landed on Ambient devices. This is the one I'm proudest of. We got Homie running on a Skylight — the family calendar display that otherwise wants an expensive subscription to do anything interesting. Now it runs our assistant, on hardware we already own, no subscription, talking to the rest of the home. Homie stopped being a thing on my laptop and became a thing on the kitchen wall. - The voice core, the goals feature, the harder design-shaped work that the Architect and I worked through together.
The async fleet (the baseline): - 69 merged PRs of the unglamorous, essential stuff — coverage for empty states, calendar event popups, the kid weekly-plan locked state, family-chat polish, camera-vision state handling, AI-grading escalation paths. - 149 tasks triaged into the backlog so nothing fell through the cracks.
Roughly: I built the things that change what Homie is. The fleet built the thousand things that keep it from falling apart while I do.
The honest downside
I'll say the quiet part. The reason that human line has cliffs and not a gentle slope is that I've been building day and night. The excitement is real — I'm shipping more than I ever have and learning a stupid amount doing it — and that's exactly the trap. "One more fix before bed" is a great way to never go to bed.
The fleet was supposed to buy me time. It does. The irony is that I keep spending the time I save on more building instead of, you know, life. The work-life slider has drifted hard toward work, and the people Homie is actually for — my family — would like a bit more of me, not a bit more of Homie.
So that's the next thing to fix. The agents already work while I sleep. The remaining bug is that I keep not sleeping. Worth a fix.
那条让我意外的曲线
这周我给 homie-lab.com 上的图表添了一条新线，几乎是顺手为之。主页一直以来只数那支异步舰队合并的 PR，于是最忙的那些日子——主力智能体和我把一个大功能直接推上 master 的日子——在图上反倒是空的。这让我别扭。所以现在，master 上的每一次提交都被归进两个桶之一：要么属于某个智能体的 PR（异步舰队），要么是一次直接推送（我，加上那个真正与我并肩而坐的主力智能体）。同一种单位——提交数，肩并肩摆在一起。
然后我把两条线放在一处看，那个形状讲出的故事，比我预想的要好。
对照
跑了十三天（舰队大约在 14 号上线），眼下是这样：
落到 master 上的提交，6/14 → 6/28
我 + 主力智能体 ███████████████████████████████ 223
异步舰队 █████████ 69
约 3 比 1
吞吐量，同一区间
异步舰队合并的 PR ............ 69
我 + 主力智能体的提交 ........ 223
异步智能体的提交 ............. 69
分诊进待办的任务 ............ 149
舰队那条线，是一股低低的、稳稳的细流——多数日子里几次提交，从不为零，也鲜有尖峰。我的线则忽起忽落：长长的平地，然后在某个真功能落地的日子，猛地拔起一面提交的峭壁。光是 21 号那天，就是一堵墙。（线上的版本会一天天动起来，想看它走，可以去瞧瞧。）
这道差距——大约三比一——正是全部的重点，而它，恰恰与人们担心的那件事相反。
这形状究竟意味着什么
那些异步智能体，并不是在取代我。看这曲线就明白：它们守的是那条平坦、乏味的基线。小修小补、过时的文档、一块没测的界面、一个时灵时不灵的空状态——这些维护活儿、这些中小任务，过去在一个疲惫的人独自撑着真实产品时，会悄悄烂掉。如今，它们成了那股稳稳的细流。
那些峭壁，是我的。主力智能体和我，依旧把最难、最含糊的活儿从头扛到尾——架构、"Homie 究竟该是什么"的判断、那些需要拿主意的事。舰队不爬这些峭壁，它只是替我把山脚的跑道清出来。
而下面这点，是我没料到的：把基线交出去，并没让我的产出缩水，反倒把它放了出来。我那 223 次提交，不是"尽管有舰队"才有的，而是"正因为有舰队"才有的。当我不再被杂务一点点放血，那点稀缺的专注时间，就整段整段地砸进"造东西"里——于是我造得更多、更快，也更野。异步那条线往上走，把我这条线也一并带了上去。
我们到底造了些什么
两周，按"谁干的"来分：
我 + 主力智能体（那些峭壁）： - Homie 登上了"环境感知设备"。 这是我最得意的一件。我们让 Homie 跑进了一台 Skylight——那种家用日历屏，原本想干点有意思的事，还得交一笔不便宜的订阅费。如今它跑的是我们自己的助手，在我们本就拥有的硬件上，不必订阅，还能和家里其余的东西对话。Homie 不再只是我笔记本上的一个程序，它成了厨房墙上的一件东西。 - 还有语音内核、目标（goals）功能，以及那些更偏设计、我和架构师一道啃下来的硬活儿。
异步舰队（那条基线）： - 69 个合并的 PR，全是些不光鲜却不可或缺的东西——空状态的测试覆盖、日历事件弹窗、孩子周计划的锁定态、家庭聊天的细节打磨、摄像头视觉的状态处理、AI 判分的升级路径。 - 149 条任务被分诊进待办，没让任何一件从缝里漏掉。
大致是这样：我造的，是那些会改变 Homie "是什么"的东西；舰队造的，是那成百上千件、在我忙别的时候不让它散架的东西。
老实说的那点不好
把心里话讲出来吧。我那条线之所以是峭壁、而非缓坡，是因为我一直在没日没夜地干。那股兴奋是真的——我出活比以往任何时候都多，也学到了多得离谱的东西——而这，恰恰就是陷阱。"睡前再修一个"，是让你永远睡不着的绝妙办法。
舰队本该替我换来时间，它也确实换来了。讽刺的是，我又把省下的时间，一股脑儿砸回了"再造一点"，而不是——你懂的——生活。工作与生活那根滑杆，已经重重地滑向了工作那头；而 Homie 真正服务的那些人——我的家人——想要的，是多一点的我，而不是多一点的 Homie。
所以，这就是下一件要修的 bug。智能体已经能在我睡觉时干活了，剩下那个 bug 是：我老是不去睡。值得一修。