Two years ago, building an admin panel for a simple website took the maintainer of Backpack for Laravel a day, maybe two. Now it takes an hour or two. He describes the shift in a detailed post, and he says the reason AI agents work so well with his package has little to do with the AI-specific features his team built.

Backpack is an admin panel package that has been around for 10 years. The maintainer says the old workflow meant installing the package, generating CRUDs, configuring every field and column, wiring up relationships and clicking through everything to check it worked. That took 30 to 60 minutes per CRUD. Now he describes what he wants, and an agent runs the package’s generators, writes the configuration and the tests, and he reviews the result. A day of work fits in an hour, he says, and a week-long feature fits in a day.

A customer from New Zealand offered a sign of the change in February 2026. The customer gave the package a 10 out of 10 in a feedback form, with a one-line review: “Straightforward to use (particularly with Claude looking over my shoulder!)”

Two popular fixes that didn’t help

The maintainer says most LLM optimization advice on making a package ready for AI starts with an llms.txt file and an MCP server. His team built both. He says neither moved the needle.

Does llms.txt help AI coding agents?

The llms.txt file, he writes, does not do much for LLM visibility. Agents rarely go looking for it, and when they do, they find a table of contents for documentation they could already read. He says he is thinking of removing it.

The MCP server was built so agents could query the documentation. Results were mixed at best. The agent has to decide to call the server, and it usually doesn’t. He also says some of the queries the AI tools sent were so weak that the server returned nothing useful. He is also considering removing it. An MCP server, he argues, suits live data an agent cannot get any other way, or tool use. Static documentation is not that kind of job.

His takeaway is simple. Anything the agent has to remember to fetch is weak. What works is information already in front of the agent, or already in its training data.

What helped by accident, and what the team did on purpose

Why do AI coding agents work well with some packages?

The biggest gains, he says, came from choices made years before anyone used AI to write code. The core API has barely changed in four years. Even version 8, which the team is working on now, started with maintenance work such as updating dependencies and fixing inconsistencies. The way a field or column is defined has stayed the same, a decision aimed at making upgrades easier for older projects. The side effect is that years of Backpack code now sit in model training data, and most of it is still correct.

The downside is that models sometimes write patterns from much older versions. The team addresses that with instructions inside each project, not by waiting for the models to catch up. Every rename, he warns, splits what a model has learned into an old way and a new way, and the model may keep suggesting the old one for years. His advice to other package authors is to deprecate rather than rename, and to keep old signatures working.

The core package, backpack/crud, is open source. That has produced a decade of GitHub issues, pull requests, tutorials, Stack Overflow answers and blog posts. Agents can search all of it when they go looking for answers.

He calls the paid code readability point the one that matters most for commercial packages. Composer installs everything into the vendor folder, including paid add-ons, with no obfuscation and no remote API. An agent can read the source of any feature and see how it works. In practice, he says, agents answer many of their own questions about Backpack by reading the vendor folder, without being told to do so.

The package has also been loose by design. Developers can override almost anything with a closure or a small tweak, without creating new classes or registering providers. A computed column, for example, takes one line: CRUD::column('full_name')->value(fn ($entry) => $entry->first_name.' '.$entry->last_name); That keeps changes in one file, which means less context for the agent to load and fewer places to go wrong.

The additions that worked

The team’s largest AI-specific win, he says, is a Laravel Boost plugin for Backpack. Developers run the install command, select backpack/crud, and their agent receives the package guidelines and skills directly in its context. Those cover which features exist, how to use them and which version-specific patterns to follow. The skills are split by topic, so an agent loads only what it needs. The plugin worked where the MCP server did not for one reason: the agent does not have to decide to fetch it.

The team also built generators. The backpack:crud command creates a model, request, controller, route and menu item in one step. A companion add-on, backpack/test-generators, writes feature tests for every CrudController, and a status command shows which controllers still lack coverage. He says a command gives the same correct result every time, while hand-written boilerplate is a fresh chance for a small error in every file.

Predictable naming also helps. A text field has a text column, and a select2 field has a select2_multiple sibling. The agent sees one pattern and guesses the rest correctly. The team also turned on a Cloudflare feature that serves Markdown instead of HTML to agents that request it, which he describes as a small but free win.

What should package authors do to help AI coding agents?

He ends with a short checklist for other package authors: put instructions in the project, stop renaming things, ship readable source, turn repeated scaffolding into commands, cut ceremony, make names guessable and skip llms.txt and docs-only MCP servers unless there is a specific reason to add them. He also urged users to adopt the best model available, saying the gap between current frontier models and those from six months ago is large, and it shows most in this kind of work. He has asked readers to share how much faster they now build admin panels.

Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.