Every application is getting its own AI assistant. A sparkle icon in the corner, a chat panel in the sidebar, a "summarize this" button.
It's the obvious move to refactor existing tools this way but I think it's usually the wrong one.
The bolted-on agent
When you build an agent into your product, you sign up for:
- Choosing the model;
- Paying for the tokens;
- Writing the prompts;
- Catching up and updating the above in an environment that is rapidly evolving.
On top of that, this agent only knows your application. It can't see the customer's codebase, their other tools, or what they were doing five minutes ago.
Meanwhile your customer probably already has an agent. Developers work with Claude Code, Codex or Cursor every day. That agent knows their project, their standards and their other tools. It's probably better than the one you'd build, and they already pay for it.
Build the interface, not the agent
The idea of Bring your own agent (BYOA) flips this around. You don't ship an agent, you ship something any agent can use well: a good CLI (commandline-interface).
Agents are very good at the terminal. Give an agent a tool with a clear --help and it'll figure out the rest, chain it with other tools, and even recover from errors.
This way your product becomes a building block in whatever workflow the user is already running, instead of another chat window they have to switch to.
GitHub is a good example. Its gh CLI lets any agent open pull requests, read issues and check builds. No built-in assistant needed: the agent you already use just picks it up.
Drupal developers already know this. Drush has been the way to script Drupal for years, and it turned out to be how an agent clears caches, runs updates, exports config and queries the database too.
Nobody had to build "Drupal AI" for that to work. Now Drupal is going a step further: since 11.4, core ships its own CLI, dr. Any module can add commands to it, with no extra dependency.
Config management and database updates follow in Drupal 12. So every Drupal site now comes with a command line an agent can use out of the box.
What that takes
Agent-friendly mostly means well-behaved. Every prompt has a flag, output comes as JSON when asked, and errors say what to try next. None of this is new. It's what good CLIs always looked like.
Where it doesn't fit
Not every user lives in a terminal, and a marketing manager won't install a CLI.
For them an embedded assistant can still make sense. But even then, build the CLI or API first and put your assistant on top of it.
You get the BYOA users for free, and your own assistant gets a clean interface to work with.
