Your Architecture Has the Wrong End User in Mind
Bloomreach's Andrew Kumar on building for agents instead of humans, why composability is now a financial argument, and what most enterprises are getting wrong about agentic infrastructure.
For the better part of two decades, Andrew Kumar has been building software for people. First websites and applications for end users, then platforms and tools for the developers building those experiences. The consumer changed once, and Kumar adapted. Then it changed again.
"For the last two decades I've been building experiences, and then the platforms and underlying tools," he says. "Now I'm working to serve the agents."
Agents don't browse interfaces, wait for pages to load, or bring implicit context to an interaction. They call APIs, chain tools, and execute tasks autonomously, often making five, six, or seven requests where a human would make one. The software infrastructure built for human users wasn't designed for this pattern of consumption, and the organizations discovering that the hard way are the ones that bolted agentic capabilities onto existing monolithic systems and assumed the work was done.
Kumar, Head of Loomi Connect at Bloomreach, now has a front-row seat to what that gap looks like in practice. And more importantly, a clear view of what closes it.
From Prototype to Production, in One Year
The year-over-year shift Kumar describes from MACH X London to MACH X Toronto is striking in its specificity. Last year, the conversation was about prototyping: what's possible, what could be built, what might work. This year, the question is what's already live, and the pressure that comes with that.
"Last year we were really prototyping the art of the possible," he says. "This year, the work is live or actively going live. There's a tremendous amount of pressure to deliver."
The competitive context makes that pressure tangible. Adobe announced MCP initiatives. Salesforce announced its own agentic layer. The platforms that enterprise technology teams have been building on for years are racing to add agent capabilities, and the organizations that treated last year's prototypes as experiments rather than foundations are finding themselves behind a curve that moved faster than anticipated.
What's changed architecturally is that the integration layer—the middleware that connects systems of record to systems of engagement—is no longer passive. It's becoming intelligent.
"REST APIs were designed for humans clicking on screens," Kumar says. "Agents don't operate that way."
At the MCP level, Bloomreach can provide richer contextual guidance so agents don't need to make repeated calls to piece together what they need. The intelligence moves earlier in the stack, and the whole system becomes faster and less expensive to run as a result.
The Financial Case for Composability
In a monolithic system, a single MCP server might expose hundreds of tools. Every agent interaction requires loading that full context, which adds latency and burns through token budgets at a rate that compounds quickly across an enterprise with multiple agents running simultaneously. "You’ll burn through it way faster," Kumar says. "And it can get very costly, very fast."
Federated MCP—the approach composable architecture enables—breaks those toolkits into intelligent groupings, limiting the context load to what's actually relevant for a given interaction. Kumar's estimate of the difference is not marginal: responses are roughly ten times faster, and token utilization is roughly ten times lower. If you’re an enterprise running AI programs at scale, that arithmetic matters enormously. It's the difference between an agentic program that compounds in value and one that gets throttled by its own infrastructure costs.
"I've arrived at this opinion not just because of my background," he says, though undoubtedly it helps. Still, he continues, "From a pure business lens, this is what makes sense."
According to Kumar, AI programs at monolithic enterprise organizations will struggle to scale because the cost and complexity will become unmanageable. The organizations that will win are the ones that made composable investments early (a conclusion that echoed across nearly every practitioner conversation from MACH X).
What Enterprises Are Getting Wrong
The most common mistake Kumar sees enterprises making isn't a technology choice. It's an assumption: that buying an agentic solution and bolting it onto existing infrastructure constitutes an agentic strategy.
"Generic, all-purpose agents are good for a basic chatbot," he says. "But anything beyond the surface starts to get specific to each business."
Every organization has processes, data structures, and workflows that are particular to how they operate. Agents that work on their behalf need to reflect that specificity, which means enterprises need to build their own agents on top of core platforms, not assume that a vendor's default configuration will do the work.
The governance implications of getting this wrong are also more acute than most enterprise teams have internalized. Human users are, by nature, a throttle on systems (there are only so many clicks a person can make in a given timeframe). Agents remove that constraint entirely.
"Agents can overwhelm systems far faster than humans," Kumar says, "which requires new controls and safeguards." Security, permissions, access controls, and rate limiting need to be designed into the agentic layer from the start, not retrofitted after the first failure.
The Democratization Nobody Expected
There's a thread running through Kumar's account of the past year that's easy to understate: the barrier to building sophisticated agentic systems has dropped so dramatically that the question of who can build them has fundamentally changed.
"There were about three of us that could do things last year," he says. "But now? Anyone could do it."
Something that required three months of specialized work and a rare combination of skills can now be achieved in hours by someone without deep technical expertise. The democratization of agent-building is real, and it's accelerating.
Alongside the change is an important architectural implication. As more people build more agents inside enterprises, the infrastructure to govern, route, and manage those agents becomes critical. Without it, the proliferation of agents creates a different kind of problem: not the inability to build, but the inability to manage what gets built. Navigating the next phase well means thinking about that governance layer now, before the volume of agents makes it unavoidable.
Leigh Bryant
Editorial Director, Composable.com
Leigh Bryant is a seasoned content and brand strategist with over a decade of experience in digital storytelling. Starting in retail before shifting to the technology space, she has spent the past ten years crafting compelling narratives as a writer, editor, and strategist.