Daily AI Builder Brief - July 4, 2026

The Operator's Take

The real shift today isn't a new model, it's a default. Claude Code v2.1.200 changed the "default" permission mode to Manual and stopped AskUserQuestion dialogs from auto-continuing. If you run unattended agents on a schedule, the harness will now stop and wait for a human instead of quietly proceeding, so a job that ran clean yesterday can silently hang tonight. Audit every background, cron, or launchd agent this week and pin --permission-mode or "defaultMode" explicitly. The pattern across this release is deliberate: the harness stopped doing things on your behalf, and that's a good thing for anyone who runs it autonomously.

Executive Summary

Claude Code default permission mode is now "Manual" across CLI, VS Code, and JetBrains (v2.1.200).
AskUserQuestion no longer auto-continues by default; idle timeout is now opt-in via /config.
Subagents cut off by rate limits or server errors now fail cleanly or return partial work instead of reporting fake success (v2.1.199 and v2.1.200).
Background sessions no longer silently stop after sleep/wake or re-run Esc-cancelled turns.
Transient 429s unrelated to your usage limit now auto-retry with backoff for subscribers.
Anthropic restored Fable 5 on July 1 behind a new cyber classifier that blocks over 99% of the reported jailbreak.
Anthropic proposed a Cyber Jailbreak Severity (CJS) framework, CJS-0 to CJS-4, to standardize how bad a jailbreak actually is.
MCP's stateless 2026-07-28 spec is in release-candidate, with beta SDKs already out ahead of the final ship.

Claude Code

v2.1.200: Manual permission mode and no more silent auto-continue

The "default" permission mode is now "Manual" everywhere, and AskUserQuestion won't auto-continue unless you opt into an idle timeout. Same release fixed background sessions stopping mid-turn after sleep/wake and daemon-handover bugs that let stale builds take over. changelog

v2.1.199: Subagents stop lying about success

Subagents cut off by a rate limit or server error used to report the failure as a successful result. They now report the error to the parent and keep partial work. TLS-proxy and SSL cert errors also fail fast with a fix hint instead of burning retries. changelog

Anthropic

Fable 5 cyber safeguards and the CJS framework

After a June incident, Anthropic restored Fable 5 on July 1 behind a classifier that blocks over 99% of the reported jailbreak, and split cyber activity into prohibited, high-risk dual-use, low-risk dual-use, and benign. It also proposed a Cyber Jailbreak Severity scale (CJS-0 to CJS-4) scoring jailbreaks on capability gain, breadth, ease of weaponization, and discoverability. Worth reading if you do agentic red teaming. detail

MCP Ecosystem

Stateless 2026-07-28 spec heading to final

No net-new MCP release inside the 48-hour window, but the shift builders should prep for is close: the 2026-07-28 spec removes Mcp-Session-Id and the session handshake, so remote servers run behind a plain load balancer with no sticky routing. Beta Python, TypeScript, Go, and C# SDKs are already out. This changes how you deploy and how you reason about MCP security. MCP blog

Broader AI

Nothing today.

What This Means For Builders

Pin --permission-mode / "defaultMode" on every unattended agent now, or expect silent stalls after the v2.1.200 default flip.
Trust subagent results again: the fixed error reporting means a "success" from a fan-out is now actually a success, so you can drop defensive re-checks you added as workarounds.
If you deploy remote MCP servers, start planning for the stateless model before July 28; sticky-session infrastructure becomes dead weight.
Adopt the CJS axes as a checklist when you evaluate model risk, even informally: capability gain, breadth, ease of weaponization, discoverability.