Solo developer behind Kubysh: 5-hour and weekly AI agent limits replaced the team lead

One person, one iPhone app and one AI agent that writes the code. That is how Kubysh, an app for personal finance and savings, is being built. The agent has two limits: a 5-hour window and a weekly cap. According to the author, over several months of work these counters have come to perform the functions of a team lead and HR staff. The post appeared on Habr on 9 October 2026. The RSS announcement shows only the intro, and the list of conclusions stays behind the cut. Below, we cover what the text says outright and what follows from it in practice.

What the text says

The headline sets the tone: when the 5-hour limit runs out, the author goes to lunch, and when the weekly limit runs out, the question is whether a day off has been declared. It is a joke, but a working routine sits behind it. The developer does not argue with the restriction or look for workarounds. Instead, he builds his day around it. The publicly available part does not say which agent or which plan is used, so we cannot give specific token consumption figures or the subscription cost here.

The "5 hours plus a week" scheme is how the subscription plans of several coding agents are structured. The short window curbs peak load, while the weekly ceiling keeps users from squeezing round-the-clock operation out of a subscription. Anthropic, for example, introduced weekly limits for Claude Code in late August 2025. A connection between that fact and the agent in the article is not confirmed; it is context only.

Why a limit works like a manager

A team lead usually does three things: prioritizes tasks, stops people from sitting too long without results, and watches for burnout. An HR manager is responsible for work routine and for making sure people go home on time. A hard resource counter performs the same functions mechanically.

• Prioritization. When the window is limited, expensive and vague tasks move to the back of the queue, and targeted fixes with a clear outcome go first.
• Stopping. An agent that hits its limit stops "tweaking just a bit more" and hands control back to the human.
• Routine. The weekly ceiling physically prevents working without days off, even if you want to.
• Planning. Consumption has to be spread across days, which means the week gets a budget, like a sprint.

This approach has a downside. A limit does not distinguish an important task from a minor one and can cut off work in the middle of a refactor. A team lead would not do that. So discipline remains the human's job: unfinished work must be saved in the repository before the counter hits zero.

Practice for those who write code with an agent

If you work alone and rely on a subscription-based agent, it makes sense to turn the limits into a planning tool rather than treat them as an annoying obstacle.

• Start the window with the heaviest task while the budget is full, and leave small fixes for the tail end.
• Keep the project specification and context in repository files, so that after a pause the agent does not spend the window getting reacquainted with the code.
• Commit before every major operation: a session cut off mid-change should not leave the working tree in a broken state.
• Use the pause between windows for reviewing diffs, manual testing on a device and writing down the next tasks. For a finance app this matters especially: the agent will not notice errors in calculations on its own.
• Split the weekly budget in advance, for example by day, so you do not burn through it before Wednesday.

The author's observation itself shows where solo development with agents is heading. The bottleneck is no longer typing speed but the quota of machine time and the human's ability to allocate it wisely. The full list of conclusions is best read in the original on Habr: judging by the announcement, it is laid out there as a set of specific correspondences between limits and managerial roles.