Claude Code's August 2026 Updates: Fork Mode, Auto Mode, More

Anthropic shipped Claude Code updates every week through August 2026, and the cumulative effect changes how a session behaves by default, not just one flashy feature. Auto mode — a classifier that decides which actions are safe to run without a prompt — became the default on Pro, Max, and Team plans starting August 14. A week later, fork mode, which lets Claude hand a side task to a background subagent that inherits the full conversation, also switched from opt-in to on-by-default in interactive sessions.

This roundup covers what shipped across five weekly digests (weeks 28, 30, 32, 33, and 34), grounded in Anthropic's own "What's New" changelog and its July 24 Claude Opus 5 announcement. If you run Claude Code daily, or you're weighing it against another CLI coding agent, the sections below cover what changed for builders, how the headline features work with real commands, and what to watch as the self-hosted and /design previews mature.

What shipped

The permission model changed first. Starting August 14, auto mode became the default permission mode for new Claude Code sessions on Pro, Max, and Team plans: a classifier evaluates each action, letting safe ones through while risky ones still get blocked or confirmed. The week before, fork mode became the default behavior in interactive sessions, spinning a side task off to a subagent that carries the full conversation instead of starting blank. Week 32 added cross-session messaging, letting sessions pass a finding to each other, plus self-hosted environments in public beta for Team and Enterprise customers.

Further back, Claude Opus 5 became the default Opus model in Claude Code on July 24, with a 1-million-token context window and a fast mode priced at $10/$50 per million tokens. Week 34 brought a research preview of /design, an artboard workflow built on Artifacts that drafts editable UI options in the CLI or Desktop app, plus a Concise output style and the ANTHROPIC_DEFAULT_MODEL environment variable for pinning the model new sessions start on. Week 28 turned /doctor into a full setup checkup, aliased /checkup, that diagnoses and fixes configuration problems.

What changed for builders

The practical effect of auto mode defaulting on is fewer interruptions — reading files, running tests, and routine edits no longer stop to ask permission, while anything flagged as risky, like a destructive shell command, still gets blocked or confirmed. Fork mode's shift to default-on changes how you delegate: instead of opening a second terminal to chase a tangent, Claude can split a side investigation into a background subagent mid-conversation while you keep working in the main thread. Both changes assume the classifier's judgment is usually right, a bigger trust step than before.

Cross-session messaging and the new @ mention syntax for naming another session address a different friction point: teams running multiple sessions on the same machine no longer have to copy-paste context between terminals. Self-hosted environments extend that further for Team and Enterprise accounts that need cloud sessions on infrastructure they control, which matters for anyone with data-residency or network-isolation requirements around agentic code execution.

How it works

The /design research preview is the clearest example of the month's push toward visual workflows. Instead of describing a UI change in prose and hoping the result matches what you pictured, /design has Claude draft multiple editable artboards — built on the same Artifacts substrate used elsewhere — and you pick the one to build from. It's still a research preview, so treat the output as a starting sketch rather than production-ready markup.

The other everyday workhorse is /doctor, now a full setup checkup rather than a quick version print, with /checkup as an alias for the same command. It walks the local environment — CLI version, auth, MCP servers, permission config — and can fix what it finds instead of just reporting it, which is the first thing worth running after any of these August changes if a session starts behaving unexpectedly.

Gotchas and compatibility

Auto mode's defaults are not a blank check: Anthropic's week 28 notes say it blocks transcript tampering outright and still asks before an rm -rf on an unresolved variable, so scripts that relied on prompt-heavy safety may need re-checking. Fork mode being on by default also means a stray side question can spin up a background subagent you didn't explicitly ask for — worth knowing before assuming every reply comes from the session you're looking at. Self-hosted environments are a public beta limited to Team and Enterprise, and GitLab merge-request support for --worktree currently covers gitlab.com URLs.

If you pin models with ANTHROPIC_DEFAULT_MODEL, check it after upgrading, since Claude Opus 5 replacing the prior default Opus model in July changed what "default" resolves to for anyone who hadn't set the variable explicitly. Opus 5's fast mode at $10/$50 per million tokens is a separate line item from the standard rate, so confirm which mode a workflow is using before assuming the fast-mode price applies.

What to watch next

The two features still labeled preview or beta are worth tracking: /design should reach general availability with a clearer artboard-to-code handoff, and self-hosted environments are still public beta for Team and Enterprise, so confirm GA terms before committing infrastructure to it. It's also worth watching whether Anthropic's Claude Security plugin, which runs a multi-agent vulnerability scan and turns findings into patches you apply yourself, gets folded into the same classifier that now governs everyday permissions.

Anthropic ships this changelog weekly at the source link below, so the fastest way to stay current is checking it directly rather than waiting for a monthly roundup like this one — features documented here as "research preview" or "public beta" are the most likely to change shape before the next update lands.