Inside Claude Code's Mid-2026 Update Wave: From Opus 4.8 to Five-Level Deep Agents

In the June 8–12 release window (v2.1.166–v2.1.176), sub-agents gained the ability to spawn their own sub-agents . Background chains are capped at five levels deep to prevent runaway concurrency, while foreground chains can spawn at any depth and are self-limiting .

A new sub-agent panel below the prompt shows the full tree with descendant counts and paths back to the main session. Running /agents opens a dashboard to observe nested work in progress .

Usability fixes: safe mode, directory switching, and model fallback

Several quality-of-life features arrived in v2.1.169:

• --safe-mode launches Claude Code with all customizations disabled—CLAUDE.md files, skills, plugins, hooks, MCP servers, and custom commands—for clean troubleshooting. Authentication, model selection, and built-in tools continue to work .
• /cd moves the active session to a different working directory without breaking the prompt cache. The new directory's CLAUDE.md is appended as a message rather than replacing the system prompt, and the session's project storage relocates so --resume and --continue find the session in the new location .
• fallbackModel allows configuring up to three fallback models tried in sequence when the primary model is overloaded or unavailable .

Auto mode expands to Pro and cloud providers

Auto mode—where Claude makes decisions on the user's behalf instead of repeatedly asking for permission—was originally launched in research preview in Week 13 (late March) . In the May 18–22 window (v2.1.143–v2.1.149), it became available on the Pro plan and gained support for Sonnet 4.6 alongside Opus .

A week later, auto mode arrived on third-party providers: Amazon Bedrock, Google Vertex AI, and Microsoft Foundry for Opus 4.7 and Opus 4.8. It replaces permission prompts with background safety checks and requires setting CLAUDE_CODE_ENABLE_AUTO_MODE=1 to opt in .

Experimental and workflow features

Several features are in research preview or have moved into general availability:

• Computer use in the CLI (Week 14, late March / early April): Claude can open native apps, click through UI elements, and verify changes directly from the terminal, closing the loop for GUI-only apps such as iOS simulators and macOS native applications. It also supports Opus 4.8 .
• /ultrareview (Week 17, late April): A public research preview that runs a fleet of bug-hunting agents in the cloud against a branch or pull request. Findings land automatically in the CLI or Desktop, and it can be pointed at a specific PR with

  /ultrareview 1234

  .
• Agent view (

  claude agents

  , Week 20, mid-May): A single screen for every Claude Code session showing what's running, blocked, or completed. It supports dispatch flags and --cwd scoping. Pressing Enter on any row opens that conversation, and pressing ← returns to the list .
• /goal (also Week 20): Sets a completion condition. Claude keeps working across turns automatically; after each turn a fast model checks the condition, and if it's not met, Claude initiates another turn. It works in interactive, -p, and Remote Control modes .
• Dynamic workflows (Week 22, late May): Claude writes an orchestration script for large-scale tasks—codebase audits, migrations—and runs it across many sub-agents in the background . The trigger keyword was later renamed to ultracode in v2.1.160 .
• /code-review (Week 22): Reports correctness bugs at a chosen effort level .

A living documentation index

To make this pace of development manageable, Anthropic introduced a weekly documentation index at code.claude.com/docs/en/whats-new. Organized by week from Week 13 through Week 24, each entry is a digest with code snippets, demos, and context on why each feature matters. The index page also links to the full changelog and model configuration docs. A machine-readable file at /docs/llms.txt lets developers fetch the complete index programmatically .