Anthropic crossed from software into physical territory this week. The Model Hardware Standard lets AI agents operate microscopes, robotic arms, and quantum lasers through a unified interface. Python SDK v1.0.0 arrived with breaking changes that will silently break CI pipelines. The MCP protocol published a roadmap that redesigns agent identity from scratch. And Claude Code shipped six releases in seven days, including a restricted mode for auditing untrusted code and patches for five security bypasses. The pre-IPO infrastructure sprint is accelerating.

Model Hardware Standard: Agents That Operate Physical Equipment

The headline story: Anthropic announced the Model Hardware Standard (MHS) on August 27 — a framework that lets AI agents discover, learn, and operate physical devices in scientific research and advanced manufacturing. Microscopes, robotic arms, liquid handlers, quantum computer lasers — all through a standardized interface.

The integration time compression is significant. What previously took weeks or months of custom integration now takes hours or minutes. The framework is model-agnostic: any agent harness can access devices via MCP, and any device with a programmable interface can participate.

Early partner results are concrete. Genentech ran a drug-discovery experiment with real-time error handling. HHMI Janelia compressed an imaging experiment from weeks to a single day. QuEra improved laser stabilization on quantum computers from 58% to 99.3% reliability.

The safety approach is deliberate. Anthropic is sharing an early version with partners to build safety evaluations before open-sourcing. A physical safety roadmap is in development. The honest acknowledgment: LLMs still lack physical intuition because they learn from text and images, not physical interaction. This is a research preview with gated access, not a production release.

For CTOs in manufacturing, pharma, or materials science: this is the first credible bridge between AI agents and your physical equipment. The MCP-based architecture means it slots into existing agent infrastructure. Watch for the open-source release.

Python SDK v1.0.0: The Migration That Bites in CI

The Anthropic Python SDK jumped to v1.0.0 on August 20. This is not a routine bump — it has breaking changes that will silently break production.

The HTTP layer moved from httpx to httpx2. Passing old httpx objects to the client throws TypeError at construction. The silent failure mode is worse: tracing, APM, and mocking tools that patch httpx will record zero SDK requests without raising errors. Your observability stack goes dark and you won’t know until something breaks upstream.

Python floor moved from 3.9 to 3.10. Legacy Text Completions API is gone entirely. temperature, top_p, and top_k were removed from messages.create(), messages.stream(), messages.parse(), and beta endpoints. AnthropicBedrock() now requires an explicit AWS region — it raises ValueError instead of defaulting to us-east-1.

Claude Code’s /claude-api upgrade python generates a migration diff for existing projects. If you can’t migrate immediately, pin anthropic==0.125.0 — the last 0.x release. But that pin is a stopgap, not a strategy.

Claude Code: Six Releases, Five Security Fixes, One Restricted Mode

The pace is extraordinary. Six versions shipped between August 22 and August 28:

v2.1.248 introduced --restricted mode (or CLAUDE_CODE_RESTRICTED=1). This creates a low-permission sandbox for auditing untrusted code or running CI. It removes built-in tools that execute commands, disables WebFetch unless explicitly named via --tools, confines file operations to the working directory, refuses bypassPermissions, and ignores all settings files. Caveat: it’s not an OS-level sandbox — credentials in environment variables are still accessible.

v2.1.251 patched five security bypasses: symlink swaps in Read/Write/Edit, plugin path traversal, beta tracing bypass, workflow tool leaks, and Grep/Glob symlink bypass. If you’re running Claude Code in production, update to v2.1.251 immediately.

Other notable additions: PreModelSwitch and PostModelSwitch hooks let admins block or confirm model switches for controlled rollouts. Live subagent streaming to Remote Control clients. Prompt cache visibility in /cost with hit ratio, misses, and warm/cold status. A spend-limit progress bar in /usage for gateway-managed deployments. /claude-api cost-optimize profiles project spend and walks through cost levers: caching, token hygiene, batching, effort level, model choice.

The bundle size dropped 14.4% (~7.8 MB). Linux x64 downloads are 4.5x smaller at ~75 MB. Native builds use 40-70 MB less memory per session. CLI startup is now non-blocking — sandbox and MCP servers no longer delay your ability to type.

MCP Roadmap: Five Priorities That Redefine Agent Infrastructure

Published August 22 by maintainers David Soria Parra and Den Delimarsky. Five priorities for the next 6-12 months, with expedited SEP review for proposals in these areas:

Agentic messaging primitives. Plain request-response doesn’t fit long-running agent loops. The roadmap matures Tasks, subscriptions/listen, and progress notifications into a coherent set. Server-initiated events via webhooks and channels eliminate polling. These will integrate into MCP core once stable.

HTTP-native transport unification. Local servers still use stdio while remote servers are HTTP-native. The plan: extend Streamable HTTP to stdio, propose HTTP/2 on standard I/O, and unify local and remote transport under a single binding.

Agent identity and security. Auth assumes a person in a browser; callers are increasingly cloud workloads. DPoP (RFC 9449) binds tokens to requesting clients so stolen bearer tokens can’t replay. Workload Identity Federation for AWS, GCP, and Azure. Long-lived pasted API keys are being designed out. Anthropic is engaging with IETF OAuth and WIMSE working groups.

Progressive discovery. Servers reveal tools gradually based on conversation instead of loading entire catalogs into context. Hundred-tool catalogs are now a known anti-pattern. Standardized tool execution result formats. ETags for caching tool and resource results.

SDK developer experience. Reference SDK generated from spec, validated against conformance test suite.

The Transports Working Group was chartered August 26. Target: 2026-12-15 spec release.

Enterprise-Managed MCP Auth Expands

Enterprise-managed authorization for MCP connectors reached GA on August 24. Admins authorize a connector once; employees inherit access automatically via identity provider groups. No consent screens, no manual OAuth flows, no per-user credentials.

New connectors this week: Datadog, Notion, Slack. Already supported: Asana, Atlassian, Canva, Figma, Granola, Linear, Supabase. Coming soon: Exa, Miro, Zoom. Okta integration via Cross App Access protocol is the launch identity provider.

Ramp provisioned approximately 2,000 employees with zero manual setup using this pattern. For enterprise admins, this is the difference between managing one connector authorization and managing 2,000 individual OAuth consents.

Community Ecosystem: 2,282 Plugins and a Discovery Problem

Anthropic launched anthropics/claude-plugins-community — a read-only GitHub mirror of security-scanned, approved community plugins. The numbers reveal a scale-versus-curation tension:

  • 286 official plugins versus 2,282 community plugins
  • Only 1.6% of community plugins list an author
  • Only 6.9% have a category
  • 99.9% have a homepage URL

The ecosystem is growing explosively but discoverability is broken. You can install plugins but finding the right one requires manual research. Expect curation layers — directories, ratings, search — to emerge as the marketplace matures.

Agent Skills became a cross-vendor standard in 2026. The SKILL.md folder format is now read by 16 agents including Codex CLI, Cursor, Copilot, Windsurf, Cline, Goose, and Kiro. Tens of thousands of public skills exist across thousands of GitHub repos. The skills repository and agentskills.io standard site launched in August.

Claude Code is now the most-used AI coding tool at work: approximately 39% of professional developers use it, and it’s the primary tool for 31%. That’s up from 18% adoption in January.

Sonnet 5 Pricing: The Clarification

Contradictory signals appeared across research sources this week. One source indicated Sonnet 5 introductory pricing would end August 31, rising from $2/$10 to $3/$15 per MTok. Enterprise research indicated the $2/$10 rate was made permanent in August, cancelling the planned increase.

The enterprise research is more consistent with Anthropic’s competitive positioning. If the increase were happening, it would be the lead story in every announcement. Treat $2/$10 as the current rate and verify directly with Anthropic before budget commits. The Batch API takes 50% off all models regardless.

Model Hardware Standard: The Strategic Signal

MHS matters beyond the technical capability. It signals Anthropic’s ambition to be the agent infrastructure layer for physical automation, not just digital workflows. The MCP-based architecture means any agent — not just Claude — can use the standard. That openness is strategically deliberate: it positions MHS as infrastructure, not a product feature.

For organizations in scientific research, pharma, or advanced manufacturing: the integration time compression from weeks to hours is transformational. But the research preview status and physical safety concerns mean this is a planning item, not a deployment item. Watch for the open-source release and safety roadmap.

Deadlines and Action Items

This week:

  • Update Claude Code to v2.1.251 — five security bypasses patched
  • Remove beta headers from production code: files-api-2025-04-14, skills-2025-10-02, ce-user-management-2026-07-13
  • Export legacy memory before September 9 if using Claude.ai memory features
  • Audit Python SDK dependencies for httpx patching — the silent breakage is the highest-risk migration item

This month:

  • Test --restricted mode for CI/CD and code auditing workflows
  • Configure enterprise-managed MCP auth via Okta for Datadog, Notion, Slack
  • Evaluate Opus 5 as default model — same price as Opus 4.8, significantly better benchmarks
  • Review MCP roadmap for architecture implications: DPoP, progressive discovery, and transport unification will change credential and tool management

Strategic:

  • If you’re in manufacturing, pharma, or materials science, evaluate MHS for your physical equipment integration roadmap
  • Budget for the MCP 2026-12-15 spec release — stateless migration, DPoP adoption, and progressive discovery will require architecture changes
  • The IPO is targeting October. Expect pricing pressure, feature acceleration, and enterprise positioning to intensify through Q4

The pattern is clear: Anthropic is building the full stack — models, developer tools, protocol standards, physical interfaces, enterprise governance, and community ecosystem — at a pace designed for public market scrutiny. The question for engineering leaders is whether your architecture can absorb the velocity.

Follow along at x.com/kkaminsk for weekly analysis of the Claude ecosystem.