• October 7, 2026
• 15 min read

Claude Code — What's New in 2026 (v2.1.117 → v2.1.293)¶

Between April 21 and October 7, 2026 Claude Code went from v2.1.116 to v2.1.293 — about 180 releases. The biggest shift is parallel work: subagents now run in the background by default, there is a screen for all your background sessions (agent view), a live tree of agents in VS Code (agent map), coordinated agent teams, script-driven workflows with dozens of agents, and cloud projects. Around that came new models (Sonnet 5 / 5.5, Opus 5 / 5.5, Fable 5.1, Haiku 5.5), auto mode as the default permission mode, mods, artifacts and a long list of smaller commands.

This page summarizes what matters for daily work. Sources: the official CHANGELOG and documentation, checked against the 2.1.293 binary.

Five Ways to Run Agents in Parallel¶

Supporting tools: worktrees (each session edits its own checkout), cross-session messaging (SendMessage / ListAgents between your sessions on any machine, since 2.1.224) and /batch (splits one large change into 5–30 worktree-isolated subagents).

How to check on running work:

Agent Map (VS Code)¶

The agent map is the VS Code extension's live view of a session's subagents (since 2.1.269).

• When a conversation has subagents, a pill such as 2 agents appears at the bottom of the prompt box. Its dot shows whether any subagent is working or waiting for your permission; it turns red after a failure.
• Click the pill to open the map: subagents drawn as a tree under the main agent, each with status, elapsed time and token count. Nested subagents sit under the agent that started them.
• Click an agent to see its prompt and tool calls, open its read-only transcript, or stop it while it runs.
• Below the agents the map lists background tasks — background shell commands and monitors — with the command's latest output on its card, refreshed while it runs (2.1.277–2.1.287).
• Type /tasks in the prompt box to open the map when no pill is showing, for example when Claude only left a dev server running.
• Run in background under a running command or subagent moves it to the background; Claude continues the turn and the item keeps its place in the map (2.1.287).
• Stop / Esc end only the current turn; background agents keep running and are stopped one by one from the map (2.1.286).

Other VS Code additions in the same period: Focus view (Ctrl+Alt+F, per-turn summaries), Hooks and Permission rules dialogs, MCP servers dialog with Add / Remove, session groups and filters, Bookmarks panel, prompt cache countdown clock, per-change Accept / Reject in diffs, /status, /sandbox, /chrome, /export, /skills, /plan dialogs.

Agent View (claude agents)¶

claude agents                                  # open agent view (full terminal)
claude --bg "investigate the flaky SettingsChangeDetector test"
claude --agent code-reviewer --bg "address review comments on PR 1234"
claude --bg --exec "npm run dev"               # a shell command as a background session
claude agents --json                           # live sessions for scripts (--all adds finished ones)
claude attach auth-fix                         # attach by (part of) the name
claude logs auth-fix                           # read its output

• Rows are grouped into Needs input / Working / Completed; pinned sessions and those waiting on you go first. Each row shows a coloured state, a headline, its age and a linked PR / MR when the session pushed one.
• Space peeks at a session and lets you reply; Enter or → attaches to the full conversation; ← from a normal session sends it to the background and opens agent view.
• Useful keys: Ctrl+T pin (pinned sessions stay alive and survive updates), Ctrl+R rename, Ctrl+F find by name, Ctrl+S group by state or directory, Ctrl+X stop (twice: delete), ? all shortcuts.
• Sessions you dispatch move into their own git worktree before editing files; when the code work is done they commit and push, and open a draft PR if the task calls for it.
• /fork copies the current conversation into a new background session with its own worktree while you keep working (since 2.1.212; the old in-session fork is now /subtask).
• defaultToAgentsView in ~/.claude.json (via /config) makes plain claude open agent view.

Subagents: Background, Forks, Nesting¶

Agent Teams¶

One session is the lead; teammates are full Claude Code sessions with their own context that share a task list and message each other — and you can talk to any teammate directly.

{
  "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }
}

• Since 2.1.178 there is no TeamCreate: each session has one implicit team, and a subagent Claude names launches as a teammate.
• Display: teammateMode — "in-process" (default), "tmux", "iterm2" or "auto".
• Teammates use the leader's model (the separate teammate-model setting was removed in 2.1.234).
• Works only in interactive sessions; in -p and the SDK named subagents stay ordinary subagents.
• Best for research and review from several angles, competing debugging hypotheses, and features split across layers. Costs noticeably more tokens than one session; partition files so teammates don't edit the same ones (teams are not worktree-isolated).

Dynamic Workflows and Ultracode¶

A workflow is a script Claude writes for a task; a runtime executes it in the background, fanning out to many subagents and cross-checking their results. Unlike a skill or an agent team, the script holds the plan, so a run is repeatable and resumable.

• Ask for one in plain words ("use a workflow to audit every endpoint for missing auth"), approve the plan, watch it with /workflows.
• /deep-research <question> is the bundled workflow: parallel web searches, cross-checked sources, a cited report.
• Save a run as a command; distribute it in a plugin.
• Ultracode: /effort ultracode (or claude --effort ultracode) makes Claude plan a workflow for every substantial task on its own. Since 2.1.284 it is a separate toggle that works at any effort level.
• Size guideline in /config (workflowSizeGuideline): small, medium (default, ~10 agents), large, unrestricted. Runs pause at a usage limit and resume when it resets.

Projects (Cloud)¶

A project is one long conversation at claude.ai/code or in the desktop app. Paste bugs, stack traces or a task list into it; Claude starts a thread (usually a cloud session) per task or routes it to the thread already working in that area. Project instructions and memory reach every thread; the Overview pane shows finished threads, PRs ready for review and threads waiting on you. A thread can also run on your own computer through Remote Control. Public beta on Pro and Max.

Code Review¶

Permissions and Safety¶

• Auto mode is the default starting mode for interactive terminal and VS Code sessions from 2.1.283, and for claude -p and the Python SDK on third-party providers from 2.1.285. A background classifier checks every action that isn't a read or an edit inside the working directory. --enable-auto-mode is gone (2.1.111); set another default with permissions.defaultMode.
• The classifier runs server-side by default on API, Enterprise, Bedrock, Vertex, Foundry and gateways (2.1.278), without billing classifier overhead; CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out where allowed.
• Auto mode refuses destructive git (reset --hard, clean -fd, stash drop, amending commits it didn't make) and terraform / pulumi / cdk destroy unless you asked for them (2.1.183); cloud metadata credential fetches and egress evasion are a separate "Containment Escape" rule (2.1.257).
• "default" mode is now labelled Manual; --permission-mode manual works (2.1.200).
• --restricted — no shell or code-running tools, no WebFetch, file tools only inside the working directory, user/project settings ignored; for evaluation harnesses on shared machines (2.1.248).
• --permission-prompts none — anything that would prompt is denied, so unattended hosts never hang (2.1.259).
• --safe-mode — start with all customizations (hooks, plugins, skills, MCP) disabled for troubleshooting (2.1.169).
• sandbox.credentials masks credential files and secret env vars from sandboxed commands (2.1.187+); sandbox.network.strictAllowlist denies unknown hosts without asking (2.1.219).
• permissions.blockReadsOutsideWorkingDirectories (2.1.257); defaultMode: "bypassPermissions" in a project file is ignored (2.1.257).

Models and Effort¶

• Effort levels low, medium, high, xhigh (plus max as a cap); /effort saves a level per model (2.1.251), s changes it for this session only.
• Advisor tool: Claude consults a stronger model at key moments (before committing to an approach, when stuck, before declaring done). /advisor opus, advisorModel setting or --advisor.
• fallbackModel takes a chain of up to three models; --fallback-model works interactively (2.1.166).
• promptCacheTtl / subagentPromptCacheTtl choose 5-minute or 1-hour caching; /cost shows prompt cache hit ratio and the likely cause of misses (2.1.251).
• Fast mode (/fast) on Opus models, about 2.5× faster at a higher price.
• TodoWrite / task tools are not offered on the newest models; CLAUDE_CODE_ENABLE_TODO_TOOLS=1 brings them back.

Mods, Artifacts, Channels and More¶

New Commands and Flags Worth Knowing¶

Breaking and Behaviour Changes¶

Check these before updating a pinned CLI in CI or a team image:

Timeline¶

See also¶

• Claude Code — Settings Reference
• Claude Code — Commands Reference
• Claude Code Hooks & Agent Personas
• Claude Code Workflow Patterns
• Claude Code Advanced Configuration
• Official: Run agents in parallel