OpenClaw 2026.11.1 adds opt-in native Agent and Workflow orchestration for Claude CLI
Operators can enable native Agent and Workflow orchestration for Claude CLI through a plugin setting that stays off by default, and Anthropic API setup can record a workspace ID.
OpenClaw 2026.11.1 lets operators turn back on Claude Code's native Agent and Workflow tools for managed Claude CLI turns, but only through a plugin setting that ships off. The setting, plugins.entries.anthropic.config.managedNativeTools, accepts only the names Agent and Workflow. Restricted runs keep their existing tool limits, so the change does not widen what a capped sender can do. The OpenClaw release notes also say saved subscription accounts can make configured models available in Claude CLI without a separate native login, though they don't explain the mechanism.
The tools went missing as a side effect of an earlier change. Pull request #183765, opened by VACInc, says that since 2026.10.7, OpenClaw-managed Claude turns have launched Claude Code with an empty tools list. That removed the native Agent (subagent) and Workflow tools with no supported way to restore them. The cause traces to #172026, which moved ordinary local turns onto the same mediated surface used for verified child completions. That surface hard-codes native tools to an empty list, and the Anthropic backend turns that into an empty --tools flag. The original change was meant to keep the tools-first cached prompt prefix from changing whenever a child completion arrived. It also removed native delegation, and no config option could select native tools on managed turns.
The pull request cites a cost gap. It attributes to the linked issue's measurements one Claude process and 341 MB for a 20-agent Workflow run, against 21 processes and 3.1 GB through sessions_spawn.
The config is a list of tool names, either as a default or per agent. An agent's own list replaces the default, and an empty list opts that agent out. The schema rejects unknown names, duplicates and unknown keys. The Anthropic backend sorts the list into a canonical order so equivalent configs produce the same command line. Only the tools list changes. The isolation flags stay in place, including the empty setting sources, the restricted settings file, disabled slash commands and strict MCP config. Skill is not offered, and Claude's own file and shell tools cannot be selected. Native subagents get the parent's tool list and act only through OpenClaw's policy-filtered MCP tools.
Restricted paths get no native tools at all. Restricted senders, delegated execution, tools.fs.workspaceOnly and caller tool caps all receive an empty list. Under exec security set to deny, or to allowlist with ask off, Agent and Workflow calls are refused before any descendant starts, so the opt-in has no effect there. In the author's live test, a capped requester with workspaceOnly enabled replied NO NATIVE AGENT TOOL and wrote no file. The uncapped run's native subagent wrote the requested file through an OpenClaw MCP tool.
The pull request also fixes a problem that it says predates the change. A warm claude-cli process serves consecutive turns, and the author found that a nested Agent could keep running after its foreground parent finished. The turn would complete and the process would stay warm. The next turn's handler and grant would then authorize the descendant's later tool calls. Claude Code 2.1.310 reparents a nested Agent once its foreground parent finishes, which hid the child from the existing bookkeeping. The new fix tracks running descendants and closes the process if any are still running when a turn ends. The stated tradeoff is that such a descendant is stopped, and the next turn cold-starts and resumes the stored Claude session.
The author's testing used real Claude Code 2.1.310 with anthropic/claude-haiku-4-5 on the claude-cli runtime, run through openclaw agent --local. The pull request says the Gateway RPC and channel path was not exercised, and that the subagent-completion and cron turns were checked in tests rather than live. It also lists documented gaps. Native calls made by descendants reach before_tool_call but do not emit after_tool_call. The linked issue measured 30 of 30 such events before the change and 0 of an unstated total afterward. A Workflow call can also read its script from a workspace file through scriptPath.
Separately, the release includes pull request #186976, by sasan1200 and merged on Nov 3, 2026. It changes Anthropic API-key login. Before the change, interactive login collected only the key, and unscoped keys then failed with HTTP 400 because no workspace header was sent. Login now asks for an optional Claude Console workspace ID, which workspace-scoped keys can leave blank. Existing workspace headers and SecretRefs are preserved. The pull request reproduced the rejection on published 2026.10.9. It reported that at the reviewed commit, the same key completed a real Gateway turn, with a relay recording HTTP 400 before the fix and HTTP 200 after.
What changed
Operators can opt in to Claude Code's native Agent and Workflow tools for managed Claude CLI turns through plugins.entries.anthropic.config.managedNativeTools, which ships off by default, after these tools were removed from managed turns since 2026.10.7.
Type: New capability · Applies to: Claude CLI, Claude Code, plugins.entries.anthropic.config.managedNativeTools, Agent, Workflow
Why it matters
Teams using Claude models through OpenClaw get an optional route to Anthropic's native orchestration, which could change how agent tasks are scheduled and permissioned.
What you can do
Keep managedNativeTools disabled until the new orchestration tools have been reviewed against the permissions your restricted runs require.
Original source
- github.comRelease openclaw 2026.11.1 · openclaw/openclaw · GitHub
- github.comfix(claude-cli): let operators keep native Agent/Workflow on managed turns by VACInc · Pull Request #183765 · openclaw/openclaw · GitHub
- github.comfix(anthropic): set API workspace at login by sasan1200 · Pull Request #186976 · openclaw/openclaw · GitHub
Where we saw it
OpenClaw releases
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works