Loopy

One enormous prompt is a bad way to run a long task. The agent loses the thread, and you lose any place to intervene before it has already committed to an approach.

Loopy makes the structure explicit. Every task becomes an ordered pipeline of blocks — small single-purpose steps like Recon, Flight Plan, Build, Red Team, Land. You describe what you want, Loopy proposes a pipeline, you reorder or drop blocks in the browser, and it executes.

What it does

• 17 built-in blocks across Understand, Design, Build, Review, and Ship
• Agent-agnostic — works with any agent CLI that runs non-interactively, defaulting to claude
• Human-in-the-loop — the pipeline pauses at review gates, where you approve or send feedback that loops back into the step
• Resumable — state is checkpointed to disk, so you can stop the server and pick up where you left off
• Browser UI — plan, edit, and watch pipelines run live over a WebSocket

Filesystem-native by choice

Loopy spawns the agent loop, the agent writes JSONL events, and a watcher picks them up. There is no bespoke IPC channel between the two processes.

That sounds primitive and is the reason it holds together: the boundary is a file format instead of a protocol, which means an agent that knows nothing about Loopy still works with it, and a crashed process leaves a readable trail rather than a lost session.