English | 中文
Design autonomous loops that can converge, prove progress, and stop.
An autonomous loop is dangerous when it keeps acting without an independent way to judge whether it is getting closer to the target. loop-builder decides whether a task deserves a loop at all, then defines the smallest governance structure that can keep the loop honest.
It does not store continuity state. ACH handles state, recovery, and handoff. loop-builder handles the semantic governance: objective acceptance, independent supervision, and stop conditions.
flowchart TD
T["Proposed long task"] --> G["Loop necessity gate"]
G -->|"not recurring / not objective / low drift risk"| N["Do not build a loop"]
G -->|"recurring + objective + drift risk"| A["Acceptance owner"]
A -->|"target and criteria"| L["Loop executor"]
L -->|"attempts and evidence"| S["ACH state root"]
S -->|"trajectory"| R["Supervisor"]
R -->|"progress is real"| L
R -->|"stalled / off target / boundary breach"| H["Stop and hand back state"]
H --> A
Build a loop only when all three are true:
- The task recurs or needs many autonomous attempts.
- Success can be judged by objective acceptance criteria.
- Drift, stalling, or self-justification is a real risk.
If any condition fails, use a normal plan instead of a loop.
| Role | Responsibility |
|---|---|
| Acceptance owner | Defines the target, criteria, and next target after success |
| Loop executor | Attempts the work and records evidence |
| Supervisor | Has authority to stop, redirect, or challenge the loop |
| ACH state root | Stores continuity state so the loop can resume without drift |
The supervisor must outrank the executor. A loop that judges itself will eventually excuse itself.
Use loop-builder. Decide whether this task should become an autonomous loop. If yes, define the acceptance criteria, executor, supervisor, stop conditions, and ACH state boundary.
Do not use it for ordinary long tasks, one-off research, vague goals, or work that cannot be objectively judged. More process does not make an unclear goal clearer.
MIT.