Claude Code’s experimental Code Mode now works with agent teams
In Claude Code version 2.1.326, the experimental Code Mode can run in the same session as agent teams. Permission checks are also tighter: in unattended runs, some commands may stop to wait for approval or be blocked.
Anthropic released version 2.1.326 of its coding agent Claude Code on GitHub on 16 November. Almost 60 per cent of the 367 changelog items are fixes and 11 are new features, by the count of the Japanese site Claude Media.
Code Mode arrived as an experimental opt-in feature in version 2.1.323, Claude Media said. In Code Mode, Claude calls tools with JavaScript it has written. Users can now turn Code Mode on while agent teams are enabled. If Code Mode is already on, sessions using teams start in it.
In agent teams, several sessions work together. Claude Media said this can easily increase token use, and the site considers it safer to limit usage before combining the features.
Claude Code also warns at startup if a value for the --allowedTools, --disallowedTools or --tools flag does not look like a complete permission rule. This lets users spot a typo that has stopped a rule from taking effect, Claude Media said. The release also fixes a bug that silently dropped a prompt typed after such a flag.
Unattended runs may stop
Many of the behaviour changes tighten permissions. In bypass permissions mode, which otherwise skips approval requests, some tool calls now require approval. If no one can respond, the call is blocked. The same applies to some Bash, PowerShell, cp and mv commands. The ClaudeKit site advises checking unattended runs after updating.
By default, sandboxed commands on macOS, Linux and WSL can no longer write to some sensitive folders in the home directory. Users can allow some of these folders with the sandbox.filesystem.allowWrite setting in user or managed settings.
In local terminal sessions, --add-dir and /add-dir now ask the user to confirm trust in a folder not yet marked as trusted. claude --bg, which runs in the background, rejects such a folder. A hook’s "decision": "block" response now takes effect in more cases, so a hook that always outputs a block decision may start blocking calls.
Some terminal and VS Code sessions now start in auto mode, in which a classifier checks Claude’s actions. This happens in organisations with the Claude Code + Cowork HIPAA configuration. It also happens in sessions with no mode set if the organisation’s policy has not loaded. In the HIPAA configuration, the disableAutoMode setting turns auto mode off.
Changes for administrators and costs
Plugins previously rejected because of a malformed display field in plugin.json now load with their hooks and MCP servers. The changelog says administrators should check the plugins they distribute.
On Linux and macOS, self-hosted runners without the gh tool now get gh api routed through Anthropic by default. It adds the machine’s GitHub sign-in, if any, to requests. The --use-anthropic-gh false flag turns it off.
Automatic compaction in Plan mode now uses the model that responds to the user, for example Opus with the opusplan setting. Compaction may therefore cost more.
Other fixes cover, for example, parallel MCP tool calls that could read another call’s result. Anthropic also fixed the claude purge command, which hid its question when output was redirected, and conversation history going missing from resumed sessions. The full list is in the GitHub release notes.
Read more: How an agent works · Agent map