The Art and Science of Loopcraft with Pi (and friends)

A talk/workshop vision for reliable, compounding agent loops.

Scope: this document governs the Loopcraft with Pi talk/workshop project in this repository. It is for talk reviewers, collaborators, future maintainers, and agents looking for product intent. It is not an agent instruction file.

This project exists to help agentic coding practitioners stop treating agents like pretend coworkers and start designing reliable, compounding loops.

Pi is the concrete surface because it makes loop state, handoffs, review, memory, and operator control visible. The idea is bigger than Pi: practitioners should leave able to use their preferred harness more deliberately by knowing when to move down for reliability and when to move up for leverage.

Who We Serve

• Primary users: software engineers already using Claude, Codex, Pi, Cursor, or similar coding agents.
• Secondary users: devtool builders, engineering leads, and workshop/conference organizers evaluating agentic coding practice.
• Not for: broad AI hype audiences, generic prompt-engineering beginners, or teams looking for a fake AI org chart.

Outcomes

• Practitioners can identify the loops inside their agent workflow instead of only writing better prompts.
• Practitioners can diagnose whether a failure needs a move down to gates/evidence or a move up to orchestration/memory/leverage.
• Practitioners can explain why model-shaped lifecycles differ from human SDLC rituals.
• The talk can become a YouTube video or workshop without losing the core spine.

Current Priorities

1. Keep the talk spine crisp: grill-with-docs as the first visible loop, then a concrete Pi coding loop, then travel through the loop stack.
2. Treat ADLC as the reliability state machine for code changes, not as the whole Loopcraft universe.
3. Pair every leverage claim with the reliability substrate that makes it safe.
4. Keep Pi central as the demo surface while making the concept harness-portable.
5. Build the slide/video surface as a Vite/Remotion app using remocn-style owned components, not a generic static deck.

Core Frame

• Open from the loop-engineering shift: “you should be designing loops that prompt your agents.” Verify exact public attribution before publishing.
• Start with grill-with-docs as the first loop: clarify what we are trying to build before code generation begins.
• Then move into a concrete coding loop visible in Pi.
• Move down into the Talk ADLC Overlay: Clarify → Rail → Build → Prosecute → Accept/Distill.
• Move up into multi-agent review, repair loops, project memory, skills, and compounding workflows.
• Use the elevator-through-the-stack visual: down = reliability, up = leverage.
• Use tldraw for diagram aesthetics and remocn/Remotion motion primitives to make the loop-stack movement visible in slides and reusable video clips.

Actors

• Audience / beneficiaries: agentic coding practitioners who want better results from real tools.
• Presenter / owner: Joel, using Pi and related harnesses as lived operating surfaces.
• Sources / influences: Latent.Space Loopcraft framing and Chris Williams’ ADLC writing.
• Not an audience: people who only want model news, vendor comparison, or beginner prompt tips.