Acquia talks how vibe coding, Hermes agents build the next content AI

Tokenmaxxing. Vibe coding. Adding structure to unstructured data. Agentic AI is not only revolutionizing content management and enterprise knowledge systems with new capabilities, but also transforming how vendors build their own cloud services.

We sat down with Acquia's Matthew Grasmick, vice president of innovation, and Vlad Pavlovic, senior director of engineering, to talk about the product roadmap of its digital experience platform, how AI influences both users and developers, and where we're all going with this.

Editor's note: This interview has been edited for clarity and brevity.

Vlad Pavlovic

How is agentic AI changing how Acquia developers work, and how they deploy new features?

Vlad Pavlovic: We're at a point where AI is here to stay and is pervasive. It's settled science. There are essentially five work surfaces where we're applying AI to the day-to-day jobs of developers: planning -- ticket refinements, story mapping, estimations, things of that nature; building -- which is how most people jumped into it at first, generating some code; verifying that what we have built is the right thing, things like code reviews and end-to-end testing; shipping code -- which is essentially automated deployment, but we're restricted in what we're allowed to ship with AI agents due to security and compliance; and operating the systems -- troubleshooting, documentation, maintenance, knowledge management.

I can confidently say we're at a point where we're no longer debating internally whether or not this is the way to go. It is more about getting better at it, getting more effective at it, and using it responsibly.

When we started our AI transformation journey, a big focus -- and a lot of companies still do it this way -- was on token spend as a way to measure effectiveness. I feel like we've evolved beyond that, where we're more focused on the efficacy of using AI to do things better and more efficiently, and less on "go out and just spend tokens and call that done."

So, you're not a big tokenmaxxing shop?

Matthew Grasmick: Personally, I get in a lot of trouble because I probably spend more tokens than anybody, and I am not even in engineering, but that's just because I've been living and breathing it for the last year and a half.

Not every team and every person in the company is at the same level of AI proficiency. There are leading teams, and there are other teams that are still catching up on understanding the best way to use these things. But we have a lot of people doing very well, definitely accelerating development and automating a lot of toil.

We're deploying a lot of Hermes agents into Slack and trying to get our product and program people filling them up with context and automating a lot of processes, so Zoom transcripts automatically get rolled up into summaries or go back into Jira -- so the conversations can actually be a way of getting things done instead of having to log into a Jira interface and update the ticket.

What are your customers telling you they need from agentic AI features and tooling?

Grasmick: It's kind of all over the place. Customers are at different points in the maturity curve, not only for agentic adoption, but also for technology in general. They work in a lot of different models -- anything from a central IT team that serves dozens or hundreds of other internal teams to a vertically integrated setup.

Some are more marketing-focused, looking at bottlenecks in conversion and campaign management. That includes things outside of digital experience -- channels like email. Others are much more developer-focused -- "Hey, we need to maintain a shared codebase that powers 1,000 sites, we need a way to easily discover vulnerabilities, patch and roll out features and do automated testing." There's a lot of variation in how it's affecting our products.

Agentic AI adoption is just beginning for most enterprises. Do you see the need for an agentic management layer -- or "control tower" as a couple of tech companies call it -- or will agents be more task-based and tactical?

Matthew Grasmick

Grasmick: It's going to become a mess if you start having hundreds of agents that you're trying to manage -- and it's going to get expensive.

What I've learned using AI myself is that the more agents you put into the mix, the messier it gets, very quickly. It's not a linear increase in complexity; it kind of scales upward. I'm using agents all the time. I love Anthropic's tools. I've used their Agent Teams features, which deploy agents in parallel. I've used their subagents, kind of a coordinating agent with many under it. I've tried having dozens of sessions in parallel, working on the same things, and -- I have to say -- it falls apart very quickly.

The best way to scale up multiple agents, is with a supervising agent. It ensures something can hold together a common thread. That is the way that we're architecting our solution -- and then it is responsible for delegating tactical tasks. It determines the scope, sends out a subagent and manages the response. There are all sorts of architectural reasons that's better -- you control what's in the context window and stay on track; you have less context drift.

You don't want hundreds of agents. You want some consolidation, and everyone's going to fight for that consolidation. But that supervisor is going to be in the customers' best interest if it's not stuck talking to a single system, such as Salesforce or Acquia.

Many SaaS vendors are introducing agentic products tailored to verticals such as telecommunications and banking. What is Acquia doing?

Grasmick: Vertical specialization is going to become more important, and we're trying to be intentional about that. We want to rapidly build governable, scalable, secure digital experience, structured data APIs -- the stuff Drupal is good at. Let's get really, really good at that, rather than make other choices.

Are Acquia customers using agents right now, or are they mostly in pilot?

Grasmick: There are, I think, 22 specialized agents that ship out of the box with Drupal CMS, which includes Drupal AI. The use cases we built represent what our customers are doing. For example, an agent that can audit content that's outdated, becoming non-compliant, or no longer getting views is very useful for customers who run a ton of [web] properties.

They're obviously also interested in the rapid build-out of new properties for microsites and campaigns, which isn't special. Still, there are lots of little vertical specialty agents that do things such as make accessibility improvements to all sites so they very quickly become compliant with standards. Or search engine optimization. This is what Drupal is good at; we specialize in content architecture or API design to ensure that it's really what you need to be multi-channel.

I don't see a ton of customers creating autonomous agents and letting them loose on something in their production applications outside of typical support bots in a chat interface, which they are doing, but that's boring. It was what we were doing a year or two ago. A lot of our customer conversations involve us trying to enable them on agentic development; a lot of them haven't started that journey.

Don Fluckinger is a seasoned B2B technology journalist with more than 30 years of experience specializing in enterprise IT, digital experience and content management. As a senior news writer at Informa TechTarget, he delivers award-winning analysis that helps IT and business leaders navigate complex technologies to enhance customer and employee experiences. Got a tip? Email him.