Spezial-Agenten

What it does

agents is a bundled React module that gives every workspace member a live overview of the configured specialized agents: which specialists exist, whether they are enabled, how many sessions are currently running, and what their recent runs produced. Admins additionally use it as the setup surface to create and edit specialists (handle, prompt, model, skills, allowed tools, auth scopes) through the embedded specialized-agent editor.

The list refreshes automatically every 15 seconds while the page is visible, so running sessions and run results stay current without manual reloads. Polling pauses while the browser tab is hidden or the module sits in a hidden kept-alive workspace tab, and resumes on the next polling tick when the workspace tab becomes visible. Returning to a hidden browser tab triggers an immediate refresh. Overview state belongs to the mounted view and is not reused across users.

How to open it

• Web: /modules/agents (bundled module, slug agents).
• Manifest: bundled-modules/agents/module.json (renderer: "react"); React UI: src/components/modules/agents-module.tsx.
• Data source: GET /api/specialized-agents/overview (agents, summary counts, and the caller's admin flag).
• The module is bundled and belongs to the minimal default installed set; admins can uninstall or reinstall it from the module store (see Configuration & limits).

Key workflows

Review specialists and activity

1. Open /modules/agents. Each row shows the specialist's avatar, name, @handle, description, enabled state, running-session count, skill and tool/auth counts, default model, core-memory setting, and last activity/run status.
2. Use the search field (matches name, handle, description, and skill labels) and the filter select (all / enabled / running / disabled) to narrow the list.
3. Expand a row (chevron button) to see its recent runs: status, invocation type (mention vs. delegation), context link, timestamps, effective model, the instruction preview, and the result or error preview.

Create or edit a specialist (admin only)

1. Admins see a + button in the header row and a pencil button per row; clicking a row also opens the editor.
2. The embedded editor (SpecializedAgentSettingsPanel) manages handle, name, description, system prompt, default model, profile image, skills, allowed tool names, auth/resource scopes, core-memory bundle, and enabled state.
3. Non-admins get a read-only overview without create/edit controls.

Self Acting Agent

Admins can opt an enabled specialist into Self Acting Agent mode in the editor. Clapilot creates a dedicated shared Aufgaben board, links it from the specialist settings, and runs a periodic specialist heartbeat using the configured interval and instructions. Each run continues watchlist work and advances open board tasks with the specialist's normal tool/auth permissions and personal memory. Approved personal memories are additionally projected into the specialist's isolated Knowledge Graph and are reachable through the knowledge_* tools only during that specialist's runs. With Scan Team Chat enabled by default, the agent also reads new messages in rooms to which it is invited, judges them against its role, and claims matching work on its own board. Each room has an independent scan cursor, the first scan looks back at most 24 hours, and the Team Chat message id is stored as the task source so repeated or overlapping agents cannot duplicate a claim. Optionally, a Team Chat status channel can be selected in the editor: the agent then posts short status updates to that channel (as itself, with its own avatar) whenever it picks up a task, finishes one, hits a blocker, or makes significant progress, giving the team live visibility into what the agent is working on. team_chat_post_message is granted implicitly for these heartbeat runs. Without a status channel, heartbeat output is never posted to chat. The per-agent graph remains invisible to users, workspace graph builds, and other specialists either way.

For an existing Self Acting Agent, the editor's Heartbeat tab shows the live heartbeat session in the same read-only conversation, activity, and details views used by Agent Orchestrator. The session appears after the agent's first heartbeat run and refreshes while the tab is open.

Use specialists from chat

Specialists are invoked from any chat by @handle mention or by delegation from the main agent; both invocation types appear in the module's run history with a link back to the originating context.

How the agent can drive it

The native runtime exposes first-class specialized-agent tools (tool-catalog family "Spezial-Agenten" / "Delegation"):

• specialized_agents_list — list configured specialists; options: include_disabled, include_prompts, limit (1-100). Admin access required.
• specialized_agents_get — load one specialist by id or handle, including prompt and scoped tool/auth configuration. Admin access required.
• specialized_agents_create — create a specialist; required: handle, name, prompt; optional: description, default_model_ref, profile_image_url, skill_keys, allowed_tool_names, allowed_auth_resource_keys, include_core_memory_tools, personal_memory_enabled, self_acting_enabled, self_acting_config, enabled, sort_order. Admin access required.
• specialized_agents_update — update an existing specialist's editable configuration by id. Admin access required.
• delegate_to_specialized_agent — delegate a focused subtask to a specialist (handle or agent_id plus message, optional base64 attachments) and return its result as a visible step in the active chat.

Channel tokens and secrets are intentionally excluded from the create/update tools; external channels are configured from the settings UI only.

Configuration & limits

• The module has no module-level settings of its own; all data comes from the specialized-agents configuration managed in /api/specialized-agents/* and the settings panel.
• Install state: the module is bundled (not fixed), so admins can uninstall it (POST /api/module-store/deactivate-bundled) or install it (POST /api/module-store/install-bundled). Instance-wide state is tracked in module_installs; the legacy disabled file is synchronized for compatibility. It does not require developer mode.
• The module's own module API (/api/modules/agents/api) is not used for data; the React UI talks to the app's specialized-agents endpoints directly.
• The page context reported to the chat agent includes total/running agent counts, active sessions, and up to 12 visible agent handles.

Troubleshooting

• No create/edit buttons: the overview endpoint returned is_admin: false; only admins can manage specialists.
• Run history looks stale: background refresh is paused while a blocking overlay/dialog is open; close it or reload the page.