Spark and Cowork — two opposite architectural bets on where the agent should live
Google's Gemini Spark is the first 24/7 cloud agent for consumers. Claude Cowork is the desktop-agent variant Anthropic moved to GA in April. Same week, opposite architectures, different customer cohorts. The split is going to define the agent product category for the next two years.
Until this month, "AI agent" in the consumer market meant a long-running session inside a user-foregrounded application: Cursor in an IDE, Claude Code in a terminal, Operator in a browser tab. The architecture was session-bounded; the agent stopped when the user closed the app. This week ended that consensus.
Google's Gemini Spark runs in the cloud and keeps working in the background when the user's phone is locked. Anthropic's Claude Cowork hit GA on April 9 as the desktop-resident agent that perceives screen pixels and drives local apps via mouse and keyboard. Two opposite architectures. Same week. Same target market.
The cloud-agent thesis
Spark is the consumer-first design. The agent runs while the user is asleep, away, in a meeting; it does work the user planned and reports back. The architectural advantage: no device dependency, no battery cost, no "close the app and lose state" constraint. The architectural disadvantage: every data input the agent uses has to be reachable from the cloud (which means in the vendor's data plane, which means data-residency and credential-handoff become first-class concerns).
The pricing logic is the AI Ultra tier. Google rationalized the Ultra pricing the same week — $99.99/mo for 5× Pro limits, $200/mo for 20× — because cloud-agent economics only work at premium subscription levels. The Goldman 24× token forecast is the demand-side argument for why; the cost curve doesn't yet allow this product at lower tiers.
The desktop-agent thesis
Cowork is enterprise-first by accident. The architecture solves the data-residency problem cloud-agents create: the agent runs on the user's machine, with the user's actual app installs, file system, and credentials. Nothing the agent touches leaves the perimeter. For regulated industries (banks, pharma, government) the desktop-agent pattern is the only deployment architecture that gets past InfoSec.
The disadvantage is the symmetric image of Spark's advantage: the agent stops when the device shuts down. There's no "work while you sleep" mode. There's also no shared-state coordination across the user's multiple devices without a separate sync layer.
How the two patterns reconcile
The likely end-state is hybrid. Managed orchestration in the cloud (Anthropic Claude Managed Agents, Google's Agent Builder, OpenAI's Operator-descendant Agent SDK) provides the trajectory planning and the long-term memory. Local execution sandboxes (Cowork-style desktop, Anthropic's self-hosted sandboxes from the May 19 release, MCP tunnels) handle the actual tool calls that need to stay inside the customer's perimeter. The user sees one product; the architecture splits across both sides.
Spark and Cowork are the visible endpoints of that hybrid. Spark is the consumer extreme (everything in the cloud). Cowork is the enterprise extreme (everything local). The mass-market enterprise products will sit between them — and the architectural decisions about what runs where are going to be the most consequential product decisions of 2026.
CNBC — Google unveils Gemini 3.5 and Gemini Spark → · TechCrunch — Google updates Gemini app at IO 2026 → · Lushbinary — AI Coding Agents 2026 Comparison →