Heidelberg AI News

Video · Cole Medin ·

11 More Tiny Coding Agent Fixes With A Stupid Amount Of Payoff (Learn These Now)

In short

The video covers 11 recurring coding-agent problems and small fixes implemented through skills, rules, hooks, and planning workflows. It matters because these changes improve reliability, validation, control, focus, and token efficiency without requiring a complete overhaul of the agent setup.

Key points

  • Use a separate reviewer agent or review skill when the agent claims unfinished work is complete.
  • Require concrete testing evidence, including commands, outputs, and application screenshots.
  • Validate applications like users with browser automation, not only unit and integration tests.
  • Convert absolute rules and workflows into hooks that can enforce them reliably.
  • Keep global rules lean and start a new conversation with a handoff document when context degrades.
  • Customize planning so humans retain key decisions and small requests stay focused.
  • Control sub-agent usage, testing depth, and tangents based on task scope and blast radius.

Written by an AI model from the video’s captions and description.

Coding agents keep hitting the same problems for you: saying the work is done when it isn't, tests that pass while the app is still broken, ignoring your rules, burning through your limits, etc.

Watch on YouTube The opening of the article at youtube.com.

In this story

A tool or model opens in the Heidelberg AI Curriculum: what it does, what it costs, how to start. A repo opens on GitHub.

Related stories

All stories

VideoCole Medin

Anthropic's New Study: You're Not Controlling AI as Much as You Think You Are

Anthropic's study examines how AI agents can quietly displace expert judgment, with analysts following the agent's recommendation 86% of the time and often repeating its reasoning as their own. For AI builders, the video argues that agents can improve productivity while still eroding independent decision-making, and recommends structuring workflows to preserve human agency.

VideoCole Medin

I REALLY Want Open Models to Win, But They Just Don't Cut It (Yet)

The video compares open-weight and frontier models on short tasks and a 12-ticket long-running AI software factory workflow. Open models are cost-effective for focused work but lose context and violate process rules over extended runs, so the recommended approach is frontier models for planning and review with open models for implementation.