Solo Developer's 'Kubysh' App: AI Agent Limits of 5 Hours and a Week Replaced the Team Lead
Kirill Popov is a solo developer building the Kubysh mobile app for iPhone. An AI agent writes the code, operating under 5-hour and weekly usage limits. These constraints have effectively taken over the role of team lead and manager: the 5-hour limit forces a lunch break with useful reflection on product goals, while the weekly limit halts all work like a mandatory day off. Popov experimented with two agents — Codex and Claude — where one could pick up where the other left off. By the end of September, he had dropped Codex, keeping only Claude as the more reliable option. • 5-hour limit = forced lunch and architecture rethink • Weekly limit = a day off that blocks releases • Used both Codex and Claude, kept only the latter
AI-processed from Habr AI; edited by Hamidun News
Kirill Popov is a solo developer behind the Kubysh mobile app for iPhone. The code is written by an AI agent operating under two constraints: a 5-hour daily usage cap and a weekly cap. Over several months of work, these limits have functionally replaced the role of a team lead and manager. Popov shared his observations on how to turn technical constraints into development management tools.
Limits as House Rules
The 5-hour limit works like a forced lunch break. When the agent exhausts its quota, it politely suggests coming back in a couple of hours. During that window, the developer can eat, go for a walk, and — most importantly — rethink architectural decisions and product goals. This turned out to be genuinely useful: the break provides the distance needed for reflection.
The weekly limit works like a mandatory day off. Work stops completely, releases get blocked, and thoughts about a pricier subscription tier start creeping in. There's logic to it: the weekly pause forces better workload distribution.
From One Agent to Two, and Back
Popov experimented with running two AI agents simultaneously: Codex and Claude. Their division of labor resembled working on a real team. If one agent hit its 5-hour limit mid-task, the other could pick up using the instruction file and work log. A concrete example from September: Codex ran out of its quota midway through a development task, and Claude read the spec file and log, then finished the job.
By the end of September, Codex was dropped. The agent started struggling with tasks and hitting its limits too quickly. Popov compares it to firing an employee: canceling a subscription equals a staff cut. Only one agent remained — Claude.
How to Structure Development Around the Limits
Popov distilled several concrete tactics:
The task lives in a file, not in the chat. The spec, execution steps, acceptance criteria, and a full work log are stored in a project file. This lets another agent (or the developer themselves, a week later) pick up the work without losing context.
Commit and push after every step. Interrupting a task mid-way costs nothing — especially if the spec is broken into small subtasks. Losing code is practically ruled out.
Heavy work goes at the start of the week. Refactoring, major architecture reviews, integration tests — all of this is best scheduled for Monday or Tuesday, while the weekly limit is still full.
Light content work toward the end of the week. By week's end, Popov saves tasks that don't require the agent: writing copy, preparing screenshots, talking to users.
On Cost and Results
The Kubysh app itself helps users track personal finances between paychecks, bringing them closer to goals like a vacation or new gadgets. Development runs on Swift and iOS. Popov invites people to try the app for free.
On the cost of Claude ($100 a month), Popov wonders whether one premium subscription is more efficient than two basic ones at $20 each. Both setups have been tried in practice. The two-agent system (Codex and Claude) showed that having an independent perspective helps when one agent can continue the other's work. But since the end of September, Claude alone has been handling it.
Question for the community: who's your team lead — the limits, or the calendar?
- Oct 27, 2026Open-source Hob brings Claude Cowork-style workflows to your laptop, no cloud required
- Oct 27, 2026Anthropic files S-1: half-year revenue of $16.3 billion, with a valuation of up to $2 trillion expected
- Oct 27, 2026Anthropic Urges Teams to Fold Subagents Back In: A Single Loop With Skills Beats Orchestrators
- Oct 27, 2026A ~/.claude Architecture With 200+ Skills: Keeping the System Prompt Under 50K Tokens and Paying $0.15 Per Session
Key Questions Answered
What are the key details on Limits as House Rules?
The 5-hour limit works like a forced lunch break. When the agent exhausts its quota, it politely suggests coming back in a couple of hours. During that window, the developer can eat, go for a walk, and — most importantly — rethink architectural decisions and product goals. This turned out to be genuinely useful: the break provides the distance needed for reflection. The weekly limit works like a mandatory day off. Work stops completely, releases get blocked, and thoughts about a pricier subscription tier start creeping in. There's logic to it: the weekly pause forces better
What are the key details on From One Agent to Two, and Back?
Popov experimented with running two AI agents simultaneously: Codex and Claude. Their division of labor resembled working on a real team. If one agent hit its 5-hour limit mid-task, the other could pick up using the instruction file and work log. A concrete example from September: Codex ran out of its quota midway through a development task, and Claude read the spec file and log, then finished the job. By the end of September, Codex was dropped. The agent started struggling with tasks and hitting its limits too quickly. Popov compares it to firing an employee: canceling a s
How to Structure Development Around the Limits?
Popov distilled several concrete tactics: The task lives in a file, not in the chat. The spec, execution steps, acceptance criteria, and a full work log are stored in a project file. This lets another agent (or the developer themselves, a week later) pick up the work without losing context. Commit and push after every step. Interrupting a task mid-way costs nothing — especially if the spec is broken into small subtasks. Losing code is practically ruled out. Heavy work goes at the start of the week. Refactoring, major architecture reviews, integration tests — all of this is
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.