Loop Engineering
Loop Engineering is a new AI programming paradigm -- designing autonomous loop systems that allow Agents to continuously execute tasks in an environment with goals, feedback, and self-verification, eliminating the need for manual prompting.
Use Cases
Use when users need to learn about "Loop Engineering", "AI loop systems", "autonomous Agent design", "/loop", "/goal", "AI programming paradigms", or "system architect".
Core Concepts
Transition from single prompt calls to continuous loop systems:
| Paradigm | How It Works | Developer Role |
|---|---|---|
| Prompt Engineering | Manually enter prompts, model executes once, human judges results | Prompt Engineer |
| Harness Engineering | Build a constrained environment for a single Agent with pre-checks/fixes/hooks, still requires human triggering | Harness Engineer |
| Loop Engineering | Manage multi-Agent sequencing, decision-making, and autonomous loops, self-driven over long periods across many rounds | System Architect / Loop Engineer |
Key Components
| Component | Purpose | Reference |
|---|---|---|
| Automations | Trigger tasks by time/event, support periodic execution and archiving | See config-examples |
| Worktrees | Manage task state and context, maintain memory and state persistence across rounds | See config-examples |
| Skills | Encapsulate reusable task capabilities as skill modules for the loop system to invoke | See patterns |
| Plugins/Connectors | Extend system capabilities, interact with external tools or services | See integration |
| Sub-agents | Small specialized Agents for division of labor, supporting parallel multi-Agent task execution | See patterns / config-examples |
| External Memory & Feedback Loops | Achieve self-assessment and correction through mechanisms like logs, tests, and type checking | See anti-patterns |
/loop and /goal Primitives
| Primitive | Behavior | Use Case | Usage Tip |
|---|---|---|---|
/loop |
Execute tasks repeatedly on a cycle | Scheduled checks, periodic maintenance | Must set maxRounds; no termination condition = infinite loop |
/goal |
Run continuously until verifiable conditions are met | Goal-oriented autonomous tasks, with an independent model checking completion each round | Completion conditions must be objectively verifiable (file exists, tests pass, metrics met) |
Quick Start
- Define the goal: Do you want scheduled patrols, or are you pursuing a verifiable objective? Use
/loopfor the former,/goalfor the latter. - Choose a pattern: Browse common loop patterns to find a design that matches your scenario.
- Configure components: Refer to configuration examples to build the loop skeleton.
- Avoid pitfalls: Read anti-patterns to prevent issues in your design early on.
- Platform integration: Refer to platform integration to connect with your actual toolchain.
Design Principles
- Goal definition: Specify verifiable completion conditions to prevent Agents from falling into meaningless loops
- Feedback mechanisms: Enable Agents to self-assess through tests, validation, logs, and other mechanisms
- State management: Use Worktrees to maintain cross-round context and memory
- Modularity: Encapsulate reusable capabilities as Skills and Plugins
- Division of labor: Reasonably partition Sub-agent responsibilities, parallelize to improve efficiency
- Graceful termination: Set timeouts, max rounds, and abnormal exit conditions
File Navigation
| File | Content | Who It's For |
|---|---|---|
| patterns.md | 8 common loop patterns + selection guide | Need to quickly find a matching pattern |
| config-examples.md | Automation / Worktree / Sub-agent configuration examples | Starting to build a loop system |
| anti-patterns.md | Common mistakes + correction strategies | Troubleshooting or preventing issues |
| integration.md | OpenClaw / TaskFlow and other platform integration | Connecting to actual toolchains |