Daily AI Builder Brief - July 27, 2026

The Operator's Take

The headline is Opus 5, but the change that reshapes your week is quieter: Claude Code now nests subagents three levels deep by default and just made Opus 5 the default model behind them. That's a real jump in how much autonomous fan-out runs before you look at it, and it lands the same week MCP's biggest rewrite since launch enters its release-candidate window. Freeze production MCP server upgrades until your SDK validates the 2026-07-28 spec, and if you run deep agent trees, cap spawn depth deliberately instead of inheriting the new default. Power you didn't ask for is still power you're paying for.

Executive Summary

Claude Opus 5 shipped July 24 as the new default Opus, with a 1M-token context window and unchanged $5/$25 per-Mtok pricing.
Opus 5 adds two reasoning-effort levels: xhigh above high, and an uncapped max tier for long-horizon agentic work.
Claude Code v2.1.219 makes Opus 5 the default and enables nested subagents to depth 3 (was 1).
v2.1.219 adds sandbox.network.strictAllowlist to deny non-allowlisted hosts without prompting.
New DirectoryAdded hook and workflowSizeGuideline settings key give more control over agent scope.
v2.1.220 followed with reliability and bug fixes.
The MCP 2026-07-28 spec release candidate is live: stateless core, sessions removed, Tasks moved to an extension.
GitHub's MCP Server already supports the next spec ahead of tomorrow's final release.

Claude Code

v2.1.219: Opus 5 default plus depth-3 subagent nesting (changelog)

Opus 5 is now the default Opus model. Subagents can spawn nested subagents up to depth 3 by default; set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to restore the old behavior. New sandbox.network.strictAllowlist denies non-allowlisted hosts for sandboxed commands without a prompt, a real hardening win for agentic runs. Also added: the DirectoryAdded hook, workflowSizeGuideline, and mcp_server_errors surfaced in the headless stream-json init event.

v2.1.220: reliability pass (changelog)

Bug fixes across terminal rendering, permission checking, MCP authentication, and remote session credential handling.

Anthropic

Claude Opus 5 (announcement, details)

Opus 5 is a step change for the Opus tier, aimed at long-running agents, coding, and computer use, at unchanged pricing. It reports 43.3% on FrontierBench v0.1 at max effort versus 18.7% for Opus 4.8. The new xhigh and max effort tiers are the practical lever: max is uncapped for long-horizon jobs. It's the default on Claude Max and strongest on Claude Pro.

MCP Ecosystem

The 2026-07-28 stateless spec (release candidate, analysis)

The largest MCP revision since launch finalizes tomorrow. It removes sessions, drops the initialization handshake, deprecates three core features, and rewrites authorization toward OAuth and OpenID Connect. Servers that needed sticky sessions can now run behind a plain round-robin load balancer. Long-running Tasks moved from core into an extension. This is breaking: the ten-week RC window is your migration runway.

GitHub MCP Server supports the next spec (changelog)

GitHub's official server already runs against the stateless spec, a useful reference implementation for testing your own migration.

Broader AI

Nothing today.

What This Means For Builders

Audit every self-hosted MCP server for session state now. If it depends on sticky sessions or the init handshake, it breaks under 2026-07-28. Test against the RC and GitHub's server before upgrading in production.
Review your subagent depth. Depth-3 nesting compounds token spend and blast radius fast; set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH to what you actually want.
Turn on sandbox.network.strictAllowlist for any agent that touches untrusted content. It closes a real exfiltration path without adding prompts.
Reserve max effort for genuinely long-horizon work. Uncapped effort means uncapped cost; default to high or xhigh and escalate on purpose.