The Operator's Take
The MCP release candidate quietly kills the protocol-level session, and that's the single change worth acting on today. If your remote MCP server relies on sticky sessions and a shared session store, it's now running on a design the spec is deprecating. The ten-week validation window is a gift, not a warning: pull the beta SDKs this week, stand your servers up behind a plain round-robin balancer, and find your stateful assumptions before the July 28 freeze forces you to. Don't wait for the final spec to start porting.
Executive Summary
- MCP 2026-07-28 release candidate is out: stateless core,
Mcp-Session-Id and protocol sessions removed, so any request can hit any instance. - Beta SDKs (Python, TypeScript, Go, C#) now support the RC; Tier 1 SDKs expected to ship inside the ten-week window.
- New MCP extensions: server-rendered UIs via MCP Apps, long-running jobs via the Tasks extension, plus OAuth/OIDC-aligned auth.
- Claude Code 2.1.198 makes subagents run in the background by default; background agents now auto-commit, push, and open draft PRs.
- Claude Code 2.1.199 hardens retries: transient 429s auto-retry with backoff for subscribers,
CLAUDE_CODE_RETRY_WATCHDOG lifts the retry cap. - Claude in Chrome is now generally available; a new
/dataviz skill ships with a runnable palette validator. - Anthropic redeployed Fable 5 globally with new classifiers and an industry jailbreak-severity scoring framework.
Claude Code
2.1.198: subagents background-by-default and auto-PR
Subagents now run in the background so the main loop keeps working, and background agents auto-commit, push, and open draft PRs on completion. Also adds Notification hook support (agent_needs_input, agent_completed) and Claude in Chrome GA. changelog
2.1.199: retry and streaming hardening
Transient rate-limit 429s (not usage limits) now auto-retry with backoff for subscribers instead of failing. Partial streaming output is preserved on mid-stream server errors, and subagents return partial work to the parent rather than failing silently. Stacked slash-skills (/skill-a /skill-b) now load up to five leading skills. changelog
Anthropic
Fable 5 redeployed with a jailbreak-severity framework
After conversations with the US government, Anthropic re-enabled Fable 5 globally with new classifiers that target and block more cybersecurity misuse, published alongside an industry framework for scoring jailbreak severity built with Amazon, Microsoft, and Google. Anthropic
MCP Ecosystem
2026-07-28 release candidate: the stateless core
The largest revision since launch. Removing the session means no sticky routing and no shared session store at the protocol layer; servers route on an Mcp-Method header and clients cache tools/list per ttlMs. Builders running agentic MCP tooling should validate against the beta SDKs now. MCP blog
Broader AI
Nothing today.
What This Means For Builders
- Port remote MCP servers to the beta SDKs during the validation window; treat any session dependency as tech debt to clear before July 28.
- Re-plan Claude Code orchestration around background-by-default subagents: expect draft PRs, and wire the new
Notification hooks so nothing completes unseen. - If rate limits were breaking long runs, 2.1.199's auto-retry and lifted cap remove a common failure; update your timeout assumptions.
Sources
1. https://code.claude.com/docs/en/changelog 2. https://www.anthropic.com/news 3. https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ 4. https://blog.modelcontextprotocol.io/posts/sdk-betas-2026-07-28/ 5. https://www.anthropic.com/news/claude-fable-5-mythos-5 |
