Google Cloud's NEXT '26 conference, held in late April 2026, centred one of its headline announcements on what the company calls Universal MCP: every Google Cloud service, from BigQuery and Cloud Storage to Spanner, is now Model Context Protocol-enabled out of the box. The stated goal is to allow AI agents to orchestrate cloud services directly, without developers needing to write custom adapters, configure credentials manually, or wire together integrations by hand.
The official keynote framed the shift in ambitious terms. "In this agentic era, you can't have humans manually filing tickets to manage capacity," Google said, describing a vision of cloud infrastructure that, in effect, drives itself.
For developers who have spent hours — sometimes entire evenings — wrestling with credential setup before writing a single line of meaningful application code, the announcement carries immediate practical appeal. Ayush Chaturvedi, a computer science student who wrote about his experience with the announcement on the DEV Community platform, described the friction as something he had previously assumed was simply part of the learning curve. "I assumed this was just the way it was," he wrote. "Turns out, professionals were also frustrated by it."
The promise of Universal MCP is that this category of overhead largely disappears. Services handle authentication and execution internally and return standardised responses through a single protocol, removing a layer of boilerplate that developers across skill levels have historically had to manage themselves.
However, developers building with Google Cloud's agent tooling report that resolving the connectivity problem exposes a deeper one. Lohith GH, another developer who documented his experience building an AI agent using Google Cloud tools after the conference, found that getting a working prototype running was relatively quick — but making it behave reliably was not.
"The agent didn't always behave predictably," he wrote, noting instances where it ignored instructions, drifted from its intended task, or produced inconsistent outputs. His conclusion: "The hardest part of building AI agents is not intelligence — it's control."
This observation points to a broader shift in what cloud development now demands of practitioners. Where traditional software engineering centres on writing explicit, deterministic logic, agent-based systems require developers to define behaviour through scope definitions, prompt design, and constraints — skills that sit outside conventional programming curricula and for which established best practices are still emerging.
Lohith identified several specific gaps in the current Google Cloud tooling: no built-in safety boundaries for agents, limited guidance on structuring long-running agentic behaviour, and difficult-to-interpret debugging when an agent makes an unexpected decision.
Security represents another open question. Chaturvedi, while broadly positive about Universal MCP, raised a pointed concern: "When everything is MCP-enabled, the attack surface grows." He noted that Google had announced security controls alongside Universal MCP but acknowledged he had not yet evaluated how robust those controls are in practice.
The picture that emerges from these developer accounts is one of genuine, meaningful progress on a longstanding pain point — paired with a set of harder, less-resolved challenges that the industry is only beginning to grapple with. Eliminating integration friction lowers the barrier to building AI agents; it does not, by itself, make those agents safe or predictable.