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:

ParadigmHow It WorksDeveloper RolePrompt EngineeringManually enter prompts, model executes once, human judges resultsPrompt EngineerHarness EngineeringBuild a constrained environment for a single Agent with pre-checks/fixes/hooks, still requires human triggeringHarness EngineerLoop EngineeringManage multi-Agent sequencing, decision-making, and autonomous loops, self-driven over long periods across many roundsSystem Architect / Loop Engineer

Key Components#

ComponentPurposeReferenceAutomationsTrigger tasks by time/event, support periodic execution and archivingSee config-examplesWorktreesManage task state and context, maintain memory and state persistence across roundsSee config-examplesSkillsEncapsulate reusable task capabilities as skill modules for the loop system to invokeSee patternsPlugins/ConnectorsExtend system capabilities, interact with external tools or servicesSee integrationSub-agentsSmall specialized Agents for division of labor, supporting parallel multi-Agent task executionSee patterns / config-examplesExternal Memory & Feedback LoopsAchieve self-assessment and correction through mechanisms like logs, tests, and type checkingSee anti-patterns

/loop and /goal Primitives#

PrimitiveBehaviorUse CaseUsage Tip/loopExecute tasks repeatedly on a cycleScheduled checks, periodic maintenanceMust set maxRounds; no termination condition = infinite loop/goalRun continuously until verifiable conditions are metGoal-oriented autonomous tasks, with an independent model checking completion each roundCompletion conditions must be objectively verifiable (file exists, tests pass, metrics met)

Quick Start#

1. Define the goal: Do you want scheduled patrols, or are you pursuing a verifiable objective? Use /loop for the former, /goal for the latter.
2. Choose a pattern: Browse common loop patterns to find a design that matches your scenario.
3. Configure components: Refer to configuration examples to build the loop skeleton.
4. Avoid pitfalls: Read anti-patterns to prevent issues in your design early on.
5. 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#

FileContentWho It's Forpatterns.md8 common loop patterns + selection guideNeed to quickly find a matching patternconfig-examples.mdAutomation / Worktree / Sub-agent configuration examplesStarting to build a loop systemanti-patterns.mdCommon mistakes + correction strategiesTroubleshooting or preventing issuesintegration.mdOpenClaw / TaskFlow and other platform integrationConnecting to actual toolchains