Codex Takes the Windows Desktop, Bedrock AgentCore SDK Ships

<p><strong>OpenAI's Codex app</strong> now controls Windows desktops — agents can see, click, and type in any live application, raising the bar for fleet sandboxing and security policy. Meanwhile, <strong>Amazon Bedrock AgentCore</strong> landed its Python SDK on PyPI with persistent S3/EFS file mounts, and the <strong>MCP spec released a stateless HTTP r...

Highlights
  • OpenAI's Codex app now supports Computer Use on Windows, letting agents see, click, and type in any desktop application — and be steered remotely from a phone while the task runs (OpenAI Developers)
  • Amazon Bedrock AgentCore's Python SDK landed on PyPI on June 2 alongside harness-level S3 and EFS file-mount support (AWS)
  • Anthropic reinstated OpenClaw and third-party agent hooks on Claude subscriptions; the June 15 Agent SDK credit-pool split remains on track and requires pre-flight budget audits now (VentureBeat)
  • The MCP specification published a release candidate targeting a fully stateless HTTP core with built-in caching headers, removing the sticky-routing requirement for horizontally scaled deployments (MCP Blog)
agent-runtime mcp billing security

Key Signals

  1. 1

    Codex Computer Use Extends to Windows Desktop

    June 1, 2026 — Codex app v26.527 brought full desktop control to Windows users: the agent can view the screen, move the pointer, and type in any foreground application while a task runs, including remote steering from iOS, Android, or a Mac. (OpenAI Developers) Sessions run on the active desktop — no background mode — and pause by default before sensitive actions such as sending email or deleting files, but that default is not a substitute for organizational sandboxing or privilege controls.

  2. 2

    Bedrock AgentCore Python SDK Published; Persistent File Mounts Land

    June 2, 2026 — The bedrock-agentcore package reached PyPI, giving teams a programmatic interface for the AgentCore harness lifecycle. AWS simultaneously added support for attaching up to five S3 or EFS access points per harness session, with the harness mounting them automatically into every session at create or update time. (AWS) This closes a gap for stateful agents that need durable storage across runs and moves AgentCore closer to infrastructure-as-code parity with Kubernetes-native deployment patterns.

  3. 3

    MCP Specification Release Candidate Drops the Session Layer

    June 2026 — The MCP spec RC removes the Mcp-Session-Id header and protocol-level session entirely, so any MCP request can land on any server replica without sticky routing or shared session stores. New Mcp-Method and Mcp-Name headers let load balancers route on operation type without body inspection, and list/resource results now carry ttlMs and cacheScope fields modeled on HTTP Cache-Control. (MCP Blog) The final spec targets July 28; teams running high-traffic MCP servers should test stateless topologies against the RC now.

Why It Matters / What To Watch

  1. Desktop Computer Use Resets the Agentic Security Bar

    • An agent that clicks and types in live Windows apps has credential-access and data-exfiltration reach comparable to a local admin session — review isolation and sandboxing policies before Codex Computer Use reaches your organization's fleet. (OpenAI Developers)
    • Codex's pause-before-sensitive-action default is a useful safety net, but mature deployments should layer policy controls — file-access scope, network egress restrictions, and approval workflows — rather than relying on model-level caution alone.
  2. Agent Infrastructure Billing Is Fragmenting Across Platforms

    • Anthropic's June 15 Agent SDK credit split, GitHub Copilot's recent AI Credits switch, and Amazon's programmatic harness SDK all signal that platform vendors are separating interactive and programmatic AI spend — teams should build cost attribution and budget alerting into agent ops pipelines now. (VentureBeat)
    • The Bedrock AgentCore Python SDK enables IaC-style harness management; AWS shops can now provision, update, and tear down agent harnesses from the same Terraform or CDK pipelines as the rest of their infrastructure. (AWS)