← all challenges

// challenge brief · hack a ton 2026

Busywork Killer

Build an AI agent that removes one painful, repetitive task from real life, with a clear before/after.

// Proposed by
OKAPI
// Industry
Productivity / automation
// Difficulty
🟢 Beginner-friendly

★ Standout builds can earn a fast-tracked interview with OKAPI.

// the problem

Everyone has that one task they dread. The founder copying numbers between spreadsheets every Monday. The freelancer re-typing invoices by hand. The NGO volunteer sorting the same inbox every morning. The sales rep pasting leads into a CRM one row at a time. It isn't hard work, it's dumb work: repetitive, manual, and quietly eating hours that should go somewhere better. Most of it was never worth building a custom tool for. An agent changes that math.

// your mission

Build an AI agent that takes over one painful, repetitive task end to end. It can be for a founder, student, freelancer, local business, NGO, city admin, sales team, operations team, or anyone who loses time to manual work. Pick a task a real person actually does, and show a clear before/after: manual chaos before, the agent quietly getting it done after.

// how it works

  1. 1Pick: choose one real, repetitive task someone does today. The narrower and more real, the better.
  2. 2Map: break the manual workflow into steps: what gets read, what gets decided, what gets produced, and where the time actually goes.
  3. 3Automate: build an agent that does those steps end to end, with a human checkpoint before anything irreversible.
  4. 4Show: demo the before/after side by side: the manual way vs. the agent, on the same real input.

// what we provide

  • This one is bring-your-own: pick a task you, your team, or someone you know actually does by hand. The realer and more specific the task, the stronger the before/after.

// suggested approach

  • Inference: for model access, we suggest [LLMok](https://llmok.app). Use code AMBASADA26 for 50% off.
  • Scope ruthlessly: one task, one person, done fully beats five tasks half-automated.
  • The win is the agent doing the work, not advising on it. Wire the real tools it needs (email, sheets, files, a form, an API) so it actually completes the job.
  • Keep a human in the loop for anything irreversible like send, pay, delete, or publish. The agent drafts, the human approves.
  • Floor version: one task, one happy path, a working demo on real input. Add edge cases and a second task from there.

// how you'll be judged

  • Is the problem real and painful? A task someone genuinely dreads, not a toy?
  • Does the demo actually work, end to end, on real input?
  • Is the before/after obvious: manual chaos vs. agent done?
  • Would someone actually use this next week?
  • Is it more than a chatbot? Does the agent take action, not just answer?

// stretch goals

  • Handle a second, related task to show the approach generalizes.
  • Learn from corrections so it gets better each run.
  • One-click setup so the person whose task it is can run it themselves.

// deliverables

  • Git repo: code + README + run instructions.
  • Demo: the before/after, live, on a real task.
  • 5-minute pitch: architecture, decisions, trade-offs.