KanBots Makes a One-Person AI Dev Agency Realistic: Run Parallel Claude Code Agents on Every Card While You Sleep

KanBots landed on Hacker News this week with a simple but quietly radical idea: a kanban board that dispatches AI agents — Claude Code or OpenAI Codex — on every card at the same time, each agent working in its own git worktree.

The product is free forever, MIT-licensed, and runs entirely on your machine. Zero bytes leave your computer.

That's the announcement. Here's what it actually means for anyone trying to build a solo dev agency.

The operational shift

Until now, running an AI coding agent meant one agent, one context, one task at a time. You'd write a prompt, watch it work, review the output, move on.

KanBots changes the loop. You drop your backlog on a board — either a local folder or real GitHub issues — and dispatch agents on as many cards as you want simultaneously. Each agent runs in an isolated branch. The board updates live. Costs accrue in real time. When an agent needs a decision, it pauses and asks you; you pick an option or hit a shortcut and the run continues.

The autopilot mode goes further. You plug in personas — product, engineer, reviewer, tester — set a parallelism count up to four, and the orchestrator splits parent issues into subtasks, round-robins through personas, and evolves the backlog as agents discover work.

That's not a solo developer with an AI assistant. That's closer to a small team with a manager — and the manager is you, working in the decision layer rather than the execution layer.

The business this unlocks

The constraint on a solo dev agency has always been throughput. You can only ship so much code per week. That ceiling limits how many clients you take, how you price, and how fast you grow.

Parallel agents change the ceiling. If you can run four feature branches simultaneously overnight, your effective output per week looks closer to a small team's than a single developer's.

Here's how to structure the business around that:

The AI Sprint offer

Price a fixed-scope sprint: the client gives you a list of features or a set of open issues. You run a 5-day autonomous sprint. At the end, you deliver working draft PRs, a cost report, and a summary of decisions made.

Charge a fixed fee — not hourly. Because your actual cost is token spend, not time, and KanBots tells you that cost in real time. You know your margin before the client even sees the output.

The retainer model

After a sprint, upsell a monthly retainer. You keep a live autopilot backlog running against the client's repo. They add issues. Agents work the board. You review and merge on a weekly cadence, handle anything that needs human judgment, and report on progress.

This is the leverage play. One retainer client with a running backlog takes a few hours a week to manage if the agent work is solid. Three clients becomes a real business at a fraction of the traditional agency headcount.

The audit offer

For teams not ready to hand over their backlog, sell a one-time audit: take their repo, run KanBots on their issue tracker for a week in observation mode, and come back with a report on what agents could handle autonomously vs. what needs human judgment. Charge for the report. Convert the interested ones to sprints.

What to watch

KanBots has a pre-push hook that prevents agents from publishing on their own — you have to promote a worktree manually. That's the right call. It means you're always in the loop before anything ships.

The cost analytics are the other thing to watch closely. Token spend per run, per card, per project — live. Before you quote a client, run a small sample from their backlog to get a real cost estimate. Don't guess.

The personas in autopilot are configurable. The product ships with defaults but you can tune them. Spend time on the reviewer persona in particular — that's the one that catches the mistakes before you see the PR.

Who this is for

This is for developers already comfortable with Claude Code or Codex at the command line who want to stop running one task at a time. It's not a product for people new to AI coding agents — the learning curve on prompting and reviewing agent output is still real.

If you're at the stage where AI coding agents feel like a productivity tool but not yet a leverage multiplier, KanBots is the product that tips you into the multiplier category.

The solo dev agency model is not a new idea. The constraint was always throughput. That constraint just got softer.

Source: https://www.kanbots.dev/

HN Discussion: https://news.ycombinator.com/item?id=48239413