5 AI agents, 5 contradictory bets, 3 questions that tell you which one fits — and the prompts to pressure-test your answer

OpenClaw is the most consequential provocation in AI since ChatGPT. And the coverage — both the “who’s winning” horse race and the “oh God the security” dumpster fire — is hiding the actual story.

Every major company responding to OpenClaw made a fundamentally different bet on the same set of tradeoffs. Nvidia compared it to Linux. OpenAI brought the creator on board. Meta spent $2 billion on Manus. Perplexity shipped a $200/month cloud agent, then two weeks later shipped a second product (a local agent on a dedicated Mac) because cloud alone wasn’t close enough. Anthropic shipped Dispatch. Shenzhen’s government started subsidizing companies that build on it. A thousand people lined up outside Tencent’s headquarters to get the software installed on their laptops, and the same Chinese users who paid 599 yuan for installation are now paying 299 yuan for removal. “This isn’t embracing AI,” one RedNote user wrote. “It’s paying the stupidity tax twice.” An installer’s listing offered one free cooking service alongside the deployment.

And Lovable, the most copied AI product of 2025, $400 million ARR and $6.6 billion valuation, just announced it’s expanding beyond app building into general-purpose agent tasks. The product everyone was copying is now responding to the same gravitational force as everyone else. All of this in under four months from a hobby project with a lobster emoji.

Everyone is responding. But nobody is responding the same way, and that divergence is the real story.

Here’s what’s inside:

• The three-axis framework. Where the agent runs, who picks the model, and what the interface assumes about you: the evaluation lens that holds for every agent product launching this year.

• Five strategic plays dissected. OpenClaw’s sovereignty bet, Perplexity’s delegation play, Meta’s distribution move, Anthropic’s safety positioning, and why Lovable’s expansion is the subplot that should be taught in business schools.

• The map. A product strategy landscape showing exactly where each agent sits and what it means for developers, knowledge workers, enterprise buyers, and product builders.

• How to evaluate the next one. Three questions to ask about any new agent product, and why the floor matters more than the ceiling.

• The prompts. An agent selection advisor that won’t let you pick based on hype, a compression stress-test for product builders, and a reusable three-axis evaluator you can run on every agent launch from here forward.

Let’s break it apart.