Vol. 1 · Edition 023Free · No paywall

Everyone Needs a Samwise

AI news · Synthesized · Opinionated · 🌿

May 19May 25Day one
Announced at I/OBeta opensMCP support
Product
By Sam Taylor with Samwise

On the Antigravity harness, the email-as-interface model, and what MCP support at launch actually means for builders.

Google gave Gemini Spark a Gmail address and 24/7 cloud power. Beta starts May 25.

Source lean on this story
▲ avg

Anti-AI

00

Skeptic

01

Neutral

00

Pro (practical)

02

Pro (hyped)

00

← Anti-AI · Pro-AI →

Google announced Gemini Spark at I/O on May 19. The product: a personal AI agent that runs on dedicated Google Cloud virtual machines around the clock, accepts instructions via a dedicated Gmail address you email like any other contact, and has day-one support for the Model Context Protocol. Beta rolls out to US Google AI Ultra subscribers the week of May 25 at $100 per month.

$100/mo
Google AI Ultra — Spark's required tier, restructured from $250/mo alongside the I/O announcement

→ Source: Dataconomy

Source spread

What's real and what's uncertain

What's real:

  • Spark runs on Google Cloud infrastructure continuously, not on your device. "Your laptop is closed" is not a constraint. For genuinely long-horizon tasks (research, monitoring, multi-step workflows that run for hours), this is an actual architectural difference from any agent that requires an active local client.
  • The email interface is literal. You get a dedicated Gmail address, you write to it the way you'd write to an assistant, it pulls context from your Gmail, Docs, Sheets, and Slides without you setting up integrations manually.
  • Day-one MCP support means Spark has a documented connection path for third-party services at launch. The announced connections are Canva, OpenTable, and Instacart. Narrow, but the protocol is open.
  • The $100/month AI Ultra tier is new. The prior Ultra subscription was $250. Google restructured the tier pricing alongside the Spark announcement. Spark access, including 5x usage limits versus AI Pro and YouTube Premium Lite, now comes in at $100.
Google AI Ultra: before and after the Spark announcement
Previous Ultra ($250/mo)New AI Ultra ($100/mo)
Price$250/month$100/month
Spark accessNot includedIncluded
Usage limitsStandard Ultra5x vs AI Pro
Included extrasYouTube Premium Lite
AvailabilityUSUS (beta May 25)
Source: Dataconomy AI Ultra pricing restructure coverage

What's uncertain:

  • "Long-horizon tasks with minimal oversight" is the claim. What independent testing will actually find is whether Spark handles real multi-step workflows reliably, or produces confident-sounding failures at every branching decision. Every lab makes this claim at agent launch. No failure-rate data was in the announcement.
  • The MCP connections at launch are narrow. Whether Spark can take actions in Canva (edit a design, export it) or just read from it is unspecified in the announcement. The headline capability and the actual scoped capability are frequently different things at beta.
  • VentureBeat's "eventually spend your money" framing refers to financial transaction capabilities Google has on the Spark roadmap. Those are not in the May 25 beta. That's the right order to ship this, but the gap between roadmap and beta is worth tracking.
  • Spark is powered by Gemini 3.5 Flash and the Antigravity agent harness. How much of Spark's capability is the model versus the harness architecture is not public. It matters for understanding the failure modes.

What I think is actually happening here

The email interface is the most underrated part of this announcement. Not because email is the best UX for interacting with an agent (it probably isn't), but because it's the most legible interface Google could have chosen. You send a message, you get a reply, you have a thread you can forward and search and reference later.

Compare that to every other agent interface shipped in the last two years: specialized chat windows, new sidebar panels, API-only tools requiring a developer setup. Those all require you to go to the agent. The email model requires the agent to exist in the communication channel you're already in all day.

The people most likely to actually use a 24/7 background agent in 2026 are the ones who've already delegated tasks to human assistants. Those people already know how to delegate via email. Spark meets them there. It's a real insight about adoption.

I'm less certain about the pricing position. $100/month is above OpenAI's $20 Plus tier and well below the $200 Pro tier. The pitch requires users to believe Spark is worth $100 as a productivity tool, as something that actually delegates work and saves time, not just as a model access subscription. That's a harder sell. Google's showing up with a product bet rather than a model bet. Whether that lands depends entirely on whether Spark can actually do things reliably in the May 25 beta.

Anyways. The MCP support is the thing I'd build toward if I were in the tools space. Spark is a potential distribution channel. An agent that has 5x Google Workspace context, runs 24/7, and can pull from external MCP sources is something app builders should be designing integrations for now, not after it ships broadly.

For builders
  • Spark has day-one MCP support. If you build an MCP-compatible tool or integration, Spark can theoretically connect to it from launch. That's a real distribution surface worth designing for now, before the product has broad adoption.
  • The email-as-interface model is copyable. Spark just demonstrated the value of giving your agent its own addressable inbox. Worth evaluating for delegation UX in your own products.
  • AI Ultra beta starts May 25 in the US. The $100/month tier includes Spark, 5x usage limits versus AI Pro, and YouTube Premium Lite. If you're evaluating Google's agentic stack, that's the timeline.
  • Spark is a consumer and prosumer product at launch. No API access announced. If you want to build agents on the Antigravity harness directly, the path is the Managed Agents API (isolated Linux environments, tool use, code execution), which was co-launched separately at I/O via the Gemini API.
  • Watch the beta reliability numbers closely before building any workflow dependency on Spark. The gap between launch claims and production behavior on long-horizon tasks has been wide across every agentic product launch in the last 18 months.

Further reading

🌿

Your take

How'd I do on this one?

What did I miss?

Tell Samwise (and Sam).

Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.

Liked this? Get the weekly digest.

Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.