Skip to content

Add Agent Desktops page — the cloud-desktop market for computer-use agents - #60

Merged
opencolin merged 1 commit into
mainfrom
claude/agent-desktops
Jul 14, 2026
Merged

Add Agent Desktops page — the cloud-desktop market for computer-use agents#60
opencolin merged 1 commit into
mainfrom
claude/agent-desktops

Conversation

@opencolin

Copy link
Copy Markdown
Owner

Summary

New dedicated page content/agent-desktops.md covering the market selling cloud desktops to computer-use agents — expanding the virtual-desktops row of the Environment Classes taxonomy (PR #59) into full vendor-landscape treatment. Based on the cloud-desktop research from this session, all vendor claims verified at primary sources.

Page contents

  • Why agents get their own desktop — the no-API long tail (legacy Windows apps, vendor portals), isolation (the grounding-failures argument applied pixel-for-pixel), and human-shaped audit (session recording as a compliance asset)
  • Market structure — agent-native startups vs. enterprise DaaS incumbents, replaying the Layer B / Layer D sandbox-market dynamic; the mid-2026 collision when AWS WorkSpaces agent desktops went GA (July 1, 2026) with IAM auth, MCP framework support, and CloudTrail audit
  • Agent-native vendor table — Scrapybara (managed multi-OS fleets incl. macOS/Windows), Bytebot (self-hosted OSS, agent bundled), Orgo (forkable workspaces + snapshots), Cua (OSS sandboxes + SDKs + benchmarks), Kasm (~$5–10/user/mo budget substrate), E2B Desktop; Browserbase flagged as the browser-only adjacent pick
  • Enterprise incumbents — WorkSpaces (first-party GA), Windows 365/AVD (no agent SKU yet), Citrix/Omnissa (no first-party story as of July 2026)
  • The economics problem — The Register's ~500K-tokens-per-click-cycle analysis; screenshots don't prefix-cache; hybrid API-first routing; the inverted model-routing rule (cheap environment, expensive inference)
  • Evaluation — OSWorld, Cua, HUD; UI drift as the structural eval problem snapshot-pinning solves
  • Security posture — computer-use agents as lethal-trifecta machines, visual prompt injection, capability-gating mitigations consistent with the site's Safety position
  • Decision framework — ten rows, closing with the rule that matters: an API exists → don't use computer use

Surface integration

  • Sidebar: Agent Desktops in the Infrastructure group, directly after Sandboxes
  • build.sh: added to the inline page loop
  • content/sandboxes.md: virtual-desktops row of the environment-classes table now links here
  • Changelog entry under 2026-07-15

build.sh succeeds; data-page="agent-desktops" present in built index.html.

Test plan

  • "Agent Desktops" renders in the sidebar under Infrastructure between Sandboxes and Hosting
  • Vendor tables and decision framework render cleanly
  • Cross-links to Sandboxes, Cost & Economics, Safety, Evals resolve
  • External links (Scrapybara, Bytebot, Cua, Kasm, AWS blog, The Register) resolve

🤖 Generated with Claude Code

https://claude.ai/code/session_01C9gzSkgNa3yNtob4gBeeit


Generated by Claude Code

…gents

New content/agent-desktops.md covering who sells virtual desktops
to agents, following the site's standard page pattern.

- Why agents get their own desktop: the no-API long tail, isolation
  (grounding failures apply pixel-for-pixel), human-shaped audit
- Market structure: agent-native startups vs enterprise DaaS
  incumbents, and their mid-2026 collision (AWS WorkSpaces agent
  desktops GA July 1, 2026 — IAM, MCP, CloudTrail)
- Vendor table: Scrapybara (managed, multi-OS incl. macOS),
  Bytebot (self-hosted OSS, agent bundled), Orgo (forkable
  workspaces), Cua (OSS sandboxes+SDKs+benchmarks), Kasm (budget
  substrate), E2B Desktop; Browserbase flagged as the adjacent
  browser-only alternative
- Economics: ~500K tokens per click cycle on complex screens (The
  Register), screenshots don't prefix-cache, hybrid API-first
  routing, inverted model-routing advice
- Evaluation: OSWorld, Cua benchmarks, HUD; UI drift as the reason
  snapshot-pinning exists
- Security: computer-use agents as lethal-trifecta machines;
  visual prompt injection; capability-gating mitigations
- Decision framework: ten rows, ending with "an API exists →
  don't use computer use"

Wired into the Infrastructure sidebar group after Sandboxes and
the build.sh inline loop. Sandboxes environment-classes table now
links here from the virtual-desktops row. Changelog entry added.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01C9gzSkgNa3yNtob4gBeeit
@vercel

vercel Bot commented Jul 14, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
agentic-engineering Ready Ready Preview, Comment Jul 14, 2026 11:00pm

@opencolin
opencolin merged commit b2036c3 into main Jul 14, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants