Tiny Spoon

⚡ Why this matters

• Cloud-hosted agent runtimes mean the developer's laptop is no longer the work environment - it's just the place the prompt gets typed. Three months ago this was Devin-only; now four vendors compete on the same primitives.
• Anthropic's $0.08/agent-hour pricing is a wedge between Claude API and the rest of the API market. It rewires the unit economics of agentic products: the platform owner takes a runtime cut on top of token revenue.
• The MCP tunnels feature (May 19) is the move regulated enterprises were waiting for - cloud agents can now reach into the customer's private VPC without exposing it. Closes the compliance objection that kept banks, hospitals, and government buyers locked into self-hosted-only stacks.

🔍 What happened

• April 8, 2026. Anthropic ships Claude Managed Agents in public beta. Composable APIs for hosted agents. gVisor-isolated containers on Anthropic infra. Sandboxed code execution, checkpointing, credential vault, scoped permissions, end-to-end tracing.
• Pricing: $0.08 per agent-runtime-hour, plus standard Claude model token usage on top.
• April 2, 2026. OpenAI Codex Cloud switches to token-based billing. Spins up a sandbox VM per task, clones the repo, returns a PR. 77.3% Terminal-Bench 2.0, ~240 tokens/sec (about 2.5x Opus throughput).
• May 7, 2026. Anthropic adds dreaming (background processing), outcomes (typed completion signals), and multiagent orchestration.
• May 19, 2026. Anthropic adds MCP tunnels (secure connection from cloud agent into the customer's private services) and self-hosted sandboxes (run the sandbox in the customer's own cloud, managed by Anthropic).
• Production customers on Anthropic stack: Notion, Rakuten, Sentry.
• Cloudflare partners with Anthropic to host Managed Agents on edge-adjacent infrastructure.
• Comparable cloud agents: OpenAI Codex Cloud, Cursor background agents (Pro+ $60/mo, Ultra $200/mo), Cognition Devin (Slack/web interface, autonomous plan-code-test-deploy loop).

💬 Smart takes

• Cloudflare blog: calls Managed Agents 'the production stack for shipping AI agents at scale' - Cloudflare is now an Anthropic hosting partner.
• InfoWorld: framed the launch as Anthropic's bid to own the agent platform layer the way AWS owns compute.
• 9to5Mac: highlighted MCP tunnels as the feature that 'lets enterprises stop running their own MCP servers.'
• Medium (unicodeveloper): warned that $0.08/hr stacks fast - long-running agents cost more than the equivalent EC2 box. Self-hosted sandboxes are the escape hatch.

🧭 Where this goes

1. OpenAI's response: a managed agent runtime under the Codex Cloud brand within 90 days. Expect tier pricing competitive with $0.08/hr.
2. Microsoft's response: Project Polaris in Copilot Studio + Azure Agent Mesh, announced at Build June 2 - already in motion.
3. Google's response: a Gemini agent runtime is the open slot. GA expected Q3 2026 if the pattern holds.
4. Enterprise IT spend shifts. A new line item appears on the FinOps dashboard - 'agent-runtime-hours' - alongside compute, storage, and tokens.
5. Cursor and Devin become acquisition targets. Whoever buys them buys the workflow layer on top of the new runtime.

🎯 Implication

• For PMs building on AI: the buy-vs-build calculus on agent infrastructure just flipped. $0.08/hr managed beats running your own gVisor sandbox stack unless you're at >1000 concurrent agents.
• For execs: the new compliance question is not 'where does the data go' but 'where does the agent EXECUTE, and what can it reach.' Map the data plane and the control plane separately.
• For developers: the IDE is becoming a thin client. If you ship code, the question 'what's my local environment' starts mattering less than 'which cloud agent runtime am I assigning the work to.'