Move Down for Reliability, Move Up for Leverage

·https://loopcraft-in-pi.wzrrd.sh/
articleaiagent-loopspiworkflowstate-machinesworkshops

maps joelclaw agent loops to an explicit down-for-gates/up-for-orchestration teaching frame

Loopcraft with Pi is the public workshop spine Joel shared for the AIE Worlds Fair: teach people using Claude, Codex, Pi, Cursor, and similar tools to stop treating agents like pretend coworkers and start designing the loops that make agent work reliable.

The sharp bit is the elevator model: move down when you need reliability, evidence, rails, review, and acceptance; move up when you want orchestration, memory, repair loops, skills, and compounding leverage. That’s a better debugging frame than “write a better prompt,” because it gives the operator an actual next move instead of more vibes.

Pi is the demo surface because it makes loop state, handoffs, review, memory, and operator control visible. The source is careful not to make this Pi-only, which matters: the pattern should survive the harness. The goal is practical agentic coding discipline, not a fake AI org chart with little robot job titles. Thank fuck.

The source also frames ADLC as a reliability state machine for code changes, not the whole universe. That distinction is useful for joelclaw because agent loops, review gates, memory, and workflow orchestration are already real system pieces here. The workshop gives those pieces a teachable shape.

Key Ideas

  • The loop is the unit of design: Loopcraft with Pi argues that practitioners should identify loops inside their agent workflow instead of only improving prompts.
  • Move down for reliability: the talk’s ADLC overlay is Clarify → Rail → Build → Prosecute → Accept/Distill, giving code changes an explicit gate-and-evidence path.
  • Move up for leverage: multi-agent review, repair loops, project memory, skills, and compounding workflows sit above the basic coding loop.
  • Pi works as the workshop surface because visible state, handoffs, review, memory, and operator control make the loop inspectable.
  • The workshop explicitly avoids broad AI hype, beginner prompt tips, and fake AI org charts; the audience is practitioners already using coding agents.
  • The planned slide/video surface uses Vite, Remotion, and tldraw aesthetics instead of a generic static deck.

Links

← All discoveries