Every cloud now sells an agent platform
the work they don't do is the work you still own

April 2026 was the month "agent platform" became a category every cloud sells. At Google Cloud Next on April 22, Google launched the Gemini Enterprise Agent Platform, positioned as the evolution of Vertex AI for building, scaling, governing, and optimizing agents. Same day, OpenAI introduced Workspace Agents in ChatGPT: shared, Codex-powered agents teams can deploy across real workflows, run on schedules, and operate inside Slack. Microsoft's Agent Framework v1.0 went GA on April 3, completing the foundational agent toolset for Azure-centric organizations.

The trade press treated it like a coronation moment for enterprise AI. The reality is more boring and more useful. Agents are now a platform primitive, the same way managed Kubernetes or managed Postgres became platform primitives a decade ago. Every cloud needs one. Every cloud sells one. The hosting question is settled.

The hosting question being settled doesn't mean the design question is settled. The platforms handle a specific set of things. They don't handle the things that actually determine whether your agent is good.

What these platforms actually do for you

They give you hosting, orchestration scaffolding, observability surfaces, tool routing, IAM hooks, and a UI to inspect runs. Some of them give you visual agent builders that non-engineers can use. All of them give you faster time-to-first-demo than rolling your own.

Those are real wins. If you've been hand-rolling orchestration glue, deployment pipelines, and observability for agent traces, the platforms remove a meaningful amount of grunt work. We're not here to argue you should reinvent any of that.

What they don't do, and why it matters

They don't define what "good" means for your workflow. They don't write your evals. They don't model your domain. They don't write the data extraction code that has to handle your specific document types. They don't decide when an agent is allowed to take a destructive action. They don't own your incident pager.

That work is still yours, and it's still where most of the cost lives. The platform is the runtime. The bespoke work is the agent.

The lock-in math is the part nobody mentions

Every one of these platforms wants you to define your tools, your prompts, your eval flows, and your orchestration in their format. That's the business model. The longer you build inside one, the harder leaving becomes.

This isn't paranoia. The cloud era is full of teams who picked a platform in year one and were paying surprise bills, fighting feature gaps, or stuck on a deprecated SDK in year three. Agent platforms will be the same. You should assume your provider will change in the next twenty-four months, either by your choice or theirs.

The mitigation is portability:

• Tool schemas and prompts live in your repo, not in the platform UI.
• Eval suites run from your CI, against any provider with a config change.
• Orchestration code is platform-aware but not platform-locked. Wrap the platform call, don't scatter it through your codebase.
• Test against at least two providers from day one, even if you only run one in production.

It's a tax. It's a smaller tax than getting trapped.

When the platform is the right call

We're not telling you to avoid these platforms. They're real platforms doing real work. There are situations where the platform path is clearly correct.

Internal tools where time-to-value beats portability. Workflows that won't outlive their first quarter. Teams without the engineering capacity to maintain a portable abstraction. Use cases where you genuinely trust the provider and the workflow won't survive a vendor switch anyway. In those cases, pick a platform and lean in.

For everything else, treat "agent platform" as a runtime choice, not as a strategy. The strategy is the agent itself, and the agent should be portable.

Why this matters for the work we care about

We've been writing about the boring triangle for months: evals, integration, and ownership. The April platform launches make that triangle more important, not less. The platforms move the work you don't have to do (provisioning, basic observability, deployment plumbing). They don't move the work that determines whether anyone can rely on what you build.

If your team is evaluating Gemini Enterprise Agent Platform, OpenAI Workspace Agents, or Microsoft Agent Framework v1.0 right now, fine. Evaluate them. Just keep the parts that actually matter to your business under your control. The platform you pick this year may not be the one you run on in two years. The agent you build is the asset that has to survive that.

Sources: Google Cloud Next agent platform coverage, April 22, 2026 (Google Developers Blog); OpenAI GPT-5.5 and workspace agents context (TechCrunch, April 23, 2026); enterprise agent platform analysis (FifthRow, April 2026); enterprise agentic AI lock-in framing (Kai Waehner, April 6, 2026).