Laravel Agent Loops

Overview

About the project

Every AI agent runs on a loop: think, act, check, repeat. That loop decides how many AI calls a task costs, how long it takes and how good the answer is. Most agent packages pick the loop for you and hide it. laravel-agent-loops makes it a choice. It sits on top of the official Laravel AI SDK, so your agents, tools and providers stay the same. Only the thinking style changes. Three loops ship out of the box: - react: answers in one go. Cheap and fast. - plan-execute: makes a plan, then works through it step by step. - reflect-retry: reviews its own answer and retries with feedback until it's approved. You can pick a loop three ways: per call with using(), per agent with a #[UseLoop] attribute, or app-wide in config. One artisan command, agent-loops:compare, runs the same task through every loop and reports steps, tokens and time. It also has an optional AI judge and JSON output for CI. In a real test with Gemini, the cheapest loop returned a list of options when the task asked for one tweet, and the self-review loop approved it. I made the reviewer stricter, and on the rerun it rejected the list and returned a single tweet. That's the point of the package: measure which loop works for your app instead of guessing. Released as v0.1.0 on Packagist. PHP 8.4, Laravel 13, MIT licensed, 46 tests with CI.

Engineering

Challenges & solutions