LoopOps Guide

June 14, 2026 ยท View on GitHub

LoopOps is the ABVX discipline for deciding when agent behavior should remain a prompt and when it should be promoted into a more durable artifact.

The core rule is simple: do not turn every repeated task into a skill. Promote only when reuse, reliability, or cost control justifies the overhead.

Promotion Ladder

  1. Prompt
    • one-off task
    • low cost of repetition
    • no recurring failure mode
  2. Checklist
    • same task repeats
    • sequence matters
    • human or agent keeps forgetting a step
  3. Skill
    • workflow is reusable and procedural
    • trigger language is stable
    • the instructions improve outcomes across tasks
  4. Script
    • a deterministic step keeps getting rewritten
    • reliability matters more than prose flexibility
  5. Loop / workflow
    • the task spans multiple cycles
    • a bounded evaluator, memory policy, and stop rule are needed

Decision Heuristics

  • Stay at the lowest artifact that reliably solves the problem.
  • Prefer a prompt when the variation surface is high and the task is still cheap.
  • Prefer a skill when the reusable value is procedural rather than code-level deterministic.
  • Prefer a script when the same code keeps being generated.
  • Prefer a loop only when bounded repetition beats manual supervision.

Anti-Patterns

  • turning one successful prompt into a published skill immediately;
  • using a skill to hide the absence of verification;
  • adding a loop without budget, stop rules, or escalation;
  • building a script when a two-line shell command already works;
  • turning local repo context into a "universal" skill without proving portability.

Example: Prompt -> Skill

Situation:

  • repeated repo tasks require the same "read narrowly, patch small, verify honestly" execution style
  • the advice is procedural, not deterministic code

Promotion:

  • prompt guidance becomes token-efficient-execution

Why not script:

  • the value is in agent behavior shaping, not a deterministic transformation

Example: Skill -> Script

Situation:

  • a skill keeps re-explaining the same JSON export, data extraction, or harness bootstrap code

Promotion:

  • keep the workflow in SKILL.md
  • move the deterministic step into scripts/

Why:

  • lower token cost
  • fewer implementation drifts

Example: Skill -> Loop

Situation:

  • repeated task flow now needs:
    • a fixed budget,
    • memory policy,
    • evaluator,
    • escalation rule,
    • bounded retries

Promotion:

  • use loopops-protocol and dynamic-workflow-packets to decide whether the work should become a supervised loop or packetized workflow
  • loopops-protocol + skillopt-evolve-skills
  • loopops-protocol + dynamic-workflow-packets
  • token-efficient-execution before promoting a workflow that may only be noisy rather than truly reusable
  • delivery-preflight-gate before longer autonomous loops

Practical Rule

If you cannot explain:

  • what recurring failure or cost this promotion fixes,
  • why the lower artifact is insufficient,
  • how the promoted artifact will be validated,

then do not promote it yet.