Collaboration

The Collaboration screen covers exactly one thing: delegation between agents — letting an agent that is in the middle of working hand a sub-task to another agent.

The everyday way in is Multi-Agent Collaboration at the bottom of the Agents page — it is the same panel. The /settings/collaboration route opens it directly too, with identical content plus a Collaboration heading.

Delegation materializes as the MCP tool croli injects into each CLI, delegate_to_agent: it is asynchronous — the parent agent hands the work off, gets a task ID back immediately, carries on with its own work, and collects the result later rather than blocking there. The switch for this tool is only here; it is not on the Extension tools screen.

General

The depth default is 1, which means "the root agent can hand work out, but the sub-agent cannot pass it further down". You only need to raise it for multi-layer orchestration; every extra layer can double the number of processes running at once.

The cache exists only while that delegating session is running, and is emptied as soon as the session ends. Over budget, the oldest results are dropped first — and what is dropped is only the parent agent's in-memory copy; the sub-agent's own session record is still there, and can still be opened and read. 0 means no limit (still cleared when the session ends).

Changes need the Save button; this section does not auto-save.

One exception notice

If a particular agent has turned on "Let the agent handle files and commands" on its own (CROLI_ACP_HOST_TOOLS=agent, set per agent on the Agents page), it does not get the delegation tools, and this section lists the names of those agents. That switch hands the handling of files and commands wholesale back to the CLI itself, and the delegation tools are withheld along with it.

Agent defaults

This tab decides what configuration the delegated call starts with, and can be overridden per agent. Which rows show up under each agent is dynamic:

• Permission mode — Croli's own set of permission presets, passed to the sub-agent as a launch parameter (not a direct mapping onto that CLI's approval policy). At the moment only Grok has this row.
• Mode — corresponds to the CLI's own notion of a mode. It only appears on its own when the agent has reported available modes and has no other configurable options — which is why something like Codex, which does have config options, never shows it.
• One row per probed option — the count is not fixed; the target agent reports them itself.

The options are not hard-coded into Croli; they come from a live probe of the target agent before launch, so what you pick is exactly what gets used. During the probe it shows "Loading available options from the agent…", and a failure can be retried; an agent that probes with nothing configurable shows "This agent has no configurable options." Every item carries an "Agent default: xxx" line, and leaving it unset follows the CLI's own default.

Worth knowing

• What gets configured here is delegation, not "several agents talking inside one conversation". Each delegation opens a separate session under the sub-agent's name, which can be opened on its own to see the whole run.
• The delegation tools are injected at the same moment as the other extension tools: when the agent starts. Changing the switch only affects agents started afterwards; a session already running is unaffected.
• Depth and cache are global, not per workspace.