Your CTO forwarded three agent platform demos this month. Anthropic on April 8. OpenAI on April 15. Google Cloud Next opening today, April 22. The ask: "Pick one by end of quarter."

Sure. And while you're at it, pick between renting an apartment, buying lumber, and pouring your own concrete foundation. Same budget line. Totally comparable.

Three launches, three different species

Between April 8 and April 22, every major AI lab shipped an "agent platform." The quotes are doing heavy lifting. Calling these the same product category is like calling a taxi service, a car dealership, and a highway toll system "transportation competitors."

Anthropic (April 8) launched Managed Agents in public beta. A managed service: you define the agent, Anthropic runs it. $0.08 per session-hour plus token fees. No infrastructure to manage. No containers to babysit. You ship a config and hope the uptime gods smile.

OpenAI (April 15) updated its Agents SDK with sandbox execution, configurable memory, and MCP (Model Context Protocol — a universal plug standard for AI tools). An open toolkit: you host it, you scale it, you debug it at 3 AM on a Saturday. Open-source. Maximum control. Maximum your-problem.

Google (April 22) opens Cloud Next with ADK v1.0, A2A protocol v0.2 (agent-to-agent communication — Google's own standard), and dual MCP/A2A support. A cloud primitive: infrastructure Lego you assemble inside GCP. First hyperscaler to natively support both agent communication protocols.

Three vendors. Three fundamentally different product categories wearing the same "agent platform" costume.

Your RFP died on arrival

Procurement teams everywhere are building evaluation matrices right now. Cute. Those matrices assume the products share comparison axes. They don't.

For a managed service, you evaluate uptime SLAs, data residency, and session pricing. For a toolkit, you evaluate developer productivity, hosting costs, and whether your team can run distributed agent infrastructure without setting the building on fire. For a cloud primitive, you evaluate protocol maturity, GCP lock-in surface, and whether A2A v0.2 survives long enough to reach v1.0.

Asking "which agent platform is best?" is like asking "what's better — Uber, a Toyota, or asphalt?" Depends on whether you have a garage, a driver, or a civil engineering degree.

The lock-in that doesn't show up in vendor slides

Here's where it gets expensive, and not in the way the pricing page warns you about.

Pick Anthropic's managed service, and you trade infrastructure headcount for vendor dependency. When the managed layer misbehaves, you file a support ticket and stare at Slack. As InfoQ reported, stealth founder Weilun Chen already flagged this: Anthropic's SDK creates lock-in, and "the trajectory definition needs to be open source." He's not wrong. You're renting someone else's abstraction, and they can remodel the apartment while you're sleeping in it.

Pick OpenAI's toolkit, and you need a platform team capable of running distributed agent infrastructure in production. That's three to five senior engineers who don't currently sit on your payroll. Your "AI initiative" just quietly became a headcount requisition with a six-month hiring timeline.

Pick Google's cloud primitives, and you're betting on two protocols — A2A at v0.2, MCP still finding its governance model — neither fully standardized. You're building load-bearing walls on a foundation that's still curing.

Each choice reshapes your engineering org in a different direction. Switching later doesn't mean rewriting a few API calls. It means restructuring teams, renegotiating contracts, and explaining to the CFO why the headcount plan changed again.

One Series C startup picked a managed agent service in March, hit a wall on custom tool orchestration within three weeks, and spent six weeks migrating to self-hosted. The code rewrite took a weekend. The organizational surgery — two infra hires, redrawn on-call rotations, rewritten SOC 2 controls — consumed the other five and a half weeks. Six weeks of zero features shipped, and a very uncomfortable board update.

What to actually do

Stop comparing platforms. Start by answering the question that actually matters: what kind of engineering org are you?

Lean team, need speed, allergic to infrastructure? Evaluate managed services. Strong platform engineers who break out in hives when they can't see the runtime? Evaluate toolkits. Already deep in GCP with multi-agent orchestration needs? Evaluate cloud primitives.

The decision isn't Anthropic vs. OpenAI vs. Google. It's managed vs. self-hosted vs. cloud-native. That choice follows from your org structure, your staffing reality, and your tolerance for 3 AM pages — not from a keynote demo.

Your CTO wanted a recommendation by end of quarter? Here it is: define your operating model first. The platform that fits will be obvious. The platform that doesn't will cost you a restructured team, six weeks of dead output, and a slide deck nobody wants to present.