× Superhuman
For the CTO22 Oct 2026Point of view · DraftDownload PDF
The connective layer · Cloudflare Developer Platform

Five products.
One brain,
at the edge.

Superhuman is now Grammarly, Mail, Docs, Go, GPTZero and, since 14 September, Fathom. Go's agents need context from all of them, in real time, per user, under tight governance. Cloudflare's Developer Platform is the connective runtime for exactly that, and no team has to re-platform to start.

01

Unified in brand. Not yet in stack.

Observed from the outside: three clouds in the request path, and no shared layer for the agents that now span them.

Edge map · observed 22 Oct 2026
Cloudflare already in the path
AWS only
Hover to trace a path
Click to jump to the play
Superhuman Go
100+ agents·~800 tools
suite site ▸ Cloudflare → CloudFront
Grammarly
40M DAU·ambient GEC
grammarly.com ▸ AWS CloudFront
Superhuman Mail
Gmail & Outlook·Mail MCP
mail.superhuman.com ▸ AWS direct
Superhuman Docs · Coda
Docs AI·Docs MCP
coda.io ▸ AWS CloudFront
Fathom
Meeting intelligence·bot-free
fathom.ai ▸ Cloudflare → Google Cloud
+ GPTZero · Calendar · Databases — same pattern, same layer
Agents SDK · Durable Objects

One stateful agent per user, mailbox, doc or meeting.

AI Gateway

One control plane for every model call.

Workers AI · Vectorize

Embeddings and inference beside the request.

Workflows · Queues

Retryable pipelines for the flaky parts.

R2

One data plane across clouds. No egress fees.

MCP Portals · Workers VPC

A governed front door, private paths back.

02

What changes in each product line

Each play names the primitive, the signal that justifies it, and the number we would measure.

Start here

Superhuman Go

Routing at 800 tools

SignalYour research: routing F1 drops 16–23% from 10 to 110 agents; embedding shortlists beat flat routing by 6–11%. Smaller, locally run embedders are on your list to explore.

  • Workers AI embeddings + Vectorize as the tool shortlister, beside the request.
  • Agents SDK + Durable Objects give every user session its own stateful agent.
  • McpAgent hosts remote MCP servers with OAuth for third-party and customer-built agents.
Prove itF1 vs. your flat-routing baseline·p95 shortlist latency

Fathom

Meetings → shared context

SignalMeeting data as agent context is the deal thesis. Fathom already sits behind Cloudflare, with a Google Cloud origin; Grammarly, Docs and Mail sit on AWS.

  • Queues + Workflows run transcribe, summarize and CRM/Slack sync, retried per step.
  • R2 for transcripts and audio, readable from AWS and GCP with no egress fees.
  • Vectorize for cross-meeting search per tenant; Deepgram models on Workers AI.
Prove itMeeting end → context in Go·pipeline recovery rate

Superhuman Mail

Mailbox state at user scale

SignalA client over Gmail and Outlook, now MCP-enabled: rate-limited, push-driven, per-account work where one slow mailbox must not touch another.

  • A Durable Object per mailbox serializes sync, holds push state and fires snooze and follow-up alarms.
  • Queues absorb provider bursts and fan out to Go agents.
  • Mail MCP on Workers with OAuth, fronted by an MCP portal.
Prove itp95 sync lag·provider 429s absorbed

Superhuman Docs

Agents as co-authors

SignalDocs MCP is live and Go lets customers add their own agents: untrusted code and tools running next to customer data.

  • Workers for Platforms gives each tenant an isolated Worker for customer-built agents.
  • Sandbox SDK + Containers for heavy tools; Browser Run for agents that act on the web.
  • Durable Objects + WebSockets for agent presence inside a live doc.
Prove itAgent cold start·blast radius of one tenant

Grammarly

Ambient AI at 100B a week

SignalAn in-house gateway hardened from ~1K to ~100K RPS across EKS + vLLM and Databricks; a silent upstream change once degraded quality mid-test.

  • Start off the hot path: AI Gateway on the long tail of new models, experiments and Go's frontier calls.
  • Workers as the front door for shadow traffic and A/B routing, the pattern you built by hand.
  • R2 as a shared eval and training data plane across AWS and Databricks.
Prove itProvider-failover time·cost per 1M long-tail calls

The whole suite

One front door, one memory

SignalMail MCP, Docs MCP and Fathom's Public API & MCP: three endpoints across AWS and GCP.

  • MCP server portals put one URL, SSO, tool-level controls and an audit log in front of all three.
  • Durable Objects + namespaced Vectorize hold permission-aware memory across mail, docs and meetings.
  • AI Search for managed RAG over whatever lands in R2.
Prove itOne endpoint, one audit trail, tool-level controls
03

Prove it in 90 days

Three bounded pilots. Each ends in a number your team already knows how to measure.

30DAYS

Shortlist bake-off

Replay Go's routing eval set against Workers AI embeddings and Vectorize. Same harness, same models, one variable changed.

Exit ▸ F1 at or above today, p95 retrieval lower
60DAYS

Stitch the context

Fathom to R2 to a Workflow to a Go context API. One MCP portal in front of Mail, Docs and Fathom.

Exit ▸ meeting to agent-ready context, measured end to end
90DAYS

Gate the models

AI Gateway on Go and experiment traffic: fallback, caching, cost tracking and spend limits. Then decide on the GEC long tail.

Exit ▸ cost and reliability read-out for your staff
The ask

Bring your routing eval set. We'll bring a Worker.

A 45-minute working session with the Go platform and Fathom infrastructure leads.

Cloudflare × SuperhumanPrepared for Superhuman engineering leadership · Draft for discussion