The Drupal and broader software community are getting overly anxious about the new kids on the block… AI. I wanted to step back and explore this anxiety through the analogy of AI as the new kids on the block, or, more specifically, the new kids entering our software teams and community. It is important to view AI not as a single kid because AI consists of multiple LLMs and harnesses.

Therefore, our immediate expectation when working with AI is that there is no single way to prepare for or interact with AI that always works across all AIs. The inconsistency and unpredictability of the current state of AI, and how it impacts our work, is making people anxious, which is leading them to want tools and processes to prepare to collaborate with AI. I'm writing this post because I think people's anxiety about AI is making them overprepare.

A large part of the AI narrative centers on the tooling you need to use AI. I feel that most of the AI tooling is overbuilt or overplanned. At the same time, harnesses like OpenCode provide essential tools and methodologies for an LLM to write code and perform tasks. Still, a harness is just a tool for AI, like the computer and software I am using to write this post.

The tooling and planning I am concerned about involve AI-specific processes for managing and orchestrating AI agents. Many people in the software community are developing AI best practices to address the challenge of integrating AI into our software development process.

A quick aside. I think having AI best practices for Drupal as a collaborative, community-led initiative is essential to Drupal's proper adoption of agent-driven development. Before someone starts using Drupal's AI best practices, they should understand what AI is.

I don't know exactly what AI is, but I've read up on some of the higher-level concepts. For me, the key point is to understand that the goal of AI is to work with humans to achieve an objective. Ultimately, an AI's interactions and outputs will be indistinguishable from those of a human, and many sci-fi novels explore this not-so-distant future. Returning to the here and now, AIs take inputs consisting of context and prompts to produce an output that can execute tools.

The biggest challenge we face with AIs and LLMs is that these systems lack a persistent or evolving memory. Literally, every request to an LLM is a blank slate that has to be reminded of what was previously requested and returned. This is the key challenge when working with an AI/LLM to build something, and my concern is that we are overthinking the solution to this, hopefully temporary, problem.

We are overthinking how to work with AI. This would be the moment to introduce my solution for working with AI, but I am not going to. I am going to offer my perspective and an analogy about the problem because I am also confused by all these solutions and feel overwhelmed by the risk of picking the wrong one. My analogy for working with AI is that AIs should be treated as the new kids on the block.

The one thing I want to keep in mind in this analogy is that the current stated limitation with AI is that every day, every interaction with AI is new, and AI has to remember and reprocess everything that happened before. First off, if a team has a new hire or is integrating a new AI, I don't think the process should be any different, because onboarding someone or something requires a process and, ideally, a plan.

Everyone has had a bad onboarding experience at a job or a bad first day of class at some point in their life. When I reflect on a bad onboarding experience, it was because the person, the leader, or the teacher wasn't welcoming, patient, or prepared. Even with AI, everyone should be patient when working with their new team member. Meanwhile, being prepared leads to a welcoming onboarding experience for someone (or something).

Creating a welcoming experience is one of the things I am most proud of in my work on the Webform module for Drupal. In the UI, when the Webform module is installed, I say, "Congratulations! You have successfully installed the Webform module. Learn more about the Webform module and Drupal." I know this message may be annoying to experienced Drupal users, but it can be dismissed or turned off. This message with links to resources and videos forced me to think about the Webform module's onboarding experience, which should be welcoming. I don't think working with AI is any different, except that an AI, with its forgetfulness, is never a prior user or contributor to your code or project.

An AI is knowledgeable about software and eager to contribute, but you have to re-onboard it to your codebase every day, more specifically at the beginning of every session. Fortunately, the concept of an AGENTS.md was created to help onboard an AI. I wish creating a single document or manual could solve our AI onboarding challenge, but it is more important to prepare for onboarding an AI as if it were its first day on the team.

We are getting to the most important takeaway from this post:

Better documentation and clearly defined roles generally lead to a more successful team and make it easier to onboard and offboard new team members. The tooling we are building for AI, including AGENTS.md, Agent skills, and specifications, aligns with the documentation and process needed for onboarding a new developer.

AGENTS.md helps orient the developer to the application's structure, with do's and don'ts. Agent Skills provide guidance for common, repeatable tasks. Specifications provide an overview of the application. Tests give a developer a sense of how the code is expected to work.

For me, everything I need to consider when working with an AI starts to align when I view AI as a new team member who needs to be onboarded daily. Yes, the daily onboarding process is frustrating, but our applications should be structured and built so that anyone can join in the morning and leave at the end of th day.

It is hard for me to give the right approach to offboarding an AI at the end of the day, or even at the end of a thread, but I think taking ownership of the code and knowing that another human, an AI, or both will look at this code tomorrow, in a month, or in a year is key. Before closing your session, ask, "Will I, someone, or something else be able to continue where I left off and move forward?"

It is ironic that ensuring your work is understandable and easy to build on is nothing new; it just requires a little extra effort, such as writing tests, which is now easier with AI.

What is really new with AI is the ability to easily do things you thought were not possible. In many ways, we have to throw out our playbook and construct a new one without ignoring all our past plays and approaches.

We don’t have to throw away everything we did before, but, sticking with the analogy that AIs are the new kids on the block, we should welcome the opportunity to change and improve how we do things. In Drupal, our code is pretty good, and we should let AI improve its quality.

With documentation, we have an opportunity to rethink how we document our code and community, making it easier for humans and AIs to contribute to Drupal. Documentation might be the single most important thing we can improve to address AI's lack of persistent memory.

Don’t overthink how to use AI. Recognize that AIs have limited memory. Accept that an AI has to be onboarded and offboarded every day. In the long term, an AI will become a highly capable team member.

For Drupal specifically, if we can improve our onboarding and offboarding for AIs, new humans will follow them. If we do this right, the new kids on the block won’t be AIs but the next generation of Drupal contributors using AI.