What Is Loopcraft?

The AI world has a new term: Loopcraft.

Some might call it a new name for an old concept. People used to call it Agent Loops or Workflow Engineering. But the shift happening in Silicon Valley is real.

Experts like Andrej Karpathy and Boris Cherny are moving away from manual prompting. They are no longer prompting agents. They are designing loops that prompt agents.

The core shift is about abstraction.

Old Way: Human -> Prompt -> Agent -> Result

New Way: Human -> Design the Loop -> Task Discovery -> Execution -> Automatic Verification -> Retry -> Save State -> Continue

Prompt Engineering optimizes one interaction. Loopcraft optimizes the entire system that runs repeatedly.

Loopcraft focuses on systemic questions:

  • Who starts the next task?
  • How does the agent know the goal?
  • Who checks the output?
  • How does failure create feedback?
  • How is state saved across sessions?

If you only automate one step, you are not building a system. You are just speeding up a manual process.

A real loop handles the entire lifecycle. For example, instead of you manually fixing a CI failure, a Loopcraft system:

  • Detects the failure.
  • Reads the logs.
  • Starts an agent in an isolated environment.
  • Modifies the code.
  • Runs tests.
  • Uses a second verifier to check the diff.
  • Creates a PR once it passes.

This is the difference between an Agent Harness and Loopcraft. A Harness is the environment. It provides tools, memory, and permissions. It answers: "How does the agent work?" Loopcraft is the orchestration. It answers: "When does the agent run, and why does it continue?"

To build a minimal loop, you need six parts:

  • Trigger: An event like a CI failure or a scheduled job.
  • Goal: A machine-verifiable condition or test.
  • State: A way to store progress and failure reasons.
  • Worker: The agent running in an isolated container.
  • Verifier: A test or a second LLM to grade the work.
  • Budget: Limits on time, cost, and retry attempts.

The highest level of Loopcraft is the improvement loop. A simple loop repeats work. An improvement loop analyzes failures to change the system itself. It adjusts prompts and tools so the next run is better.

Stop trying to fix every single result manually. Start fixing the system that produces those results.

Source: https://dev.to/luhuidev/what-is-loopcraft-from-prompt-engineering-to-agent-loop-system-design-2dff

Optional learning community: https://t.me/GyaanSetuAi