Agents

Agents are the primary unit of execution in SwarmClaw. Each agent has a personality, runtime configuration, tool selection, memory defaults, optional MCP servers, and delegation settings.

Creating an Agent

Go to Agents and create a new agent.

The default visible setup path is intentionally short:

  1. Basics: name, avatar, description
  2. Model & Connection: provider, model, credentials, optional OpenClaw route
  3. Instructions: soul/personality plus system prompt
  4. Behavior: heartbeat toggle
  5. Advanced Settings: memory, voice, delegation, MCP, routing, limits, and runtime controls

Defaults for New Agents

New agents start with safer, higher-autonomy defaults:

  • Heartbeat: on
  • Proactive memory recall: on
  • Conversation skill drafting: on
  • Delegation target mode: All Agents when delegation is enabled

Existing agents keep their saved settings when edited.

Soul vs System Prompt

  • Soul / Personality defines how the agent sounds and behaves.
  • System Prompt defines technical behavior, workflow expectations, and hard constraints.

SwarmClaw prepends user preferences and runtime context before the system prompt so agents still know the current date, chat state, and operating constraints.

Tools, MCP Servers, and Extensions

SwarmClaw separates three capability layers:

  • Tools: built-in capability families such as execute, files, web, browser, memory, manage_platform, or manage_connectors
  • MCP servers: external Model Context Protocol tool sources injected into selected agents
  • Extensions: external .js or .mjs add-ons installed from the Extensions surface

Canonical persisted fields are:

  • tools
  • extensions
  • mcpServerIds

Delegation and Orchestrator Mode

Delegation and orchestrator mode are separate controls:

  • Delegation lets an agent hand work to other agents or external CLI runtimes.
  • Orchestrator mode lets an eligible agent wake on a schedule, review platform state, and decide whether to delegate, notify, or take other operator actions.

Delegation is controlled by:

  • Can Delegate to Other Agents
  • Allowed Delegate Agents

Target selection supports:

  • All Agents
  • Selected Agents

Orchestrator mode is configured per agent with:

  • Orchestrator Mode
  • Mission
  • Wake Interval
  • Governance
  • Max Cycles / Day

Orchestrator mode is available only for eligible non-OpenClaw, non-worker-only providers. If you switch an agent onto an ineligible provider, SwarmClaw disables orchestrator mode automatically.

See Delegation for the exact delegation model.

OpenClaw Routing

Any agent can use an OpenClaw gateway route. When OpenClaw mode is enabled, the agent still behaves like a normal SwarmClaw agent, but execution is routed through the selected OpenClaw gateway/profile.

Gateway preferences matter only for OpenClaw-routed agents.

Memory Defaults

Agents can choose memory behavior in advanced settings:

  • memory scope preference
  • proactive recall behavior
  • Standard or Strict planning behavior
  • session reset policy
  • identity continuity behavior

External connector sender sessions are automatically forced onto session-scoped memory so one external contact cannot poison another sender's memory context.

Planning Mode

Agent advanced settings include a planning mode selector:

  • Standard: no extra plan contract; the agent can answer, plan, or act normally based on the task.
  • Strict planning: before multi-step tool work, the agent must emit SwarmClaw's machine-readable plan block so the runtime can track the current step and completed steps.

Strict planning is useful for coordinator agents, release agents, or any agent that regularly performs multi-step work where operators need a visible execution plan.

Configuration History

Editing an agent records a prior-version snapshot. Existing agents show Configuration History in Advanced Settings with recent saved versions and restore controls.

Restoring a version reloads the agent record and closes the sheet so the next edit starts from the restored settings.

Voice and Runtime Skills

Agents can optionally set a voice override and pin managed skills. They can also draft new skills from conversations and use discovered skills at runtime.

Conversation skill drafting is enabled by default for new agents, but can be disabled per agent.

Agent Chats and Thread Types

Each agent can have:

  • a main chat or thread
  • isolated direct connector sessions in Inbox
  • delegated task or structured-session threads

The connector owner's own conversation can stay on the main agent thread, while other external senders stay isolated in Inbox sessions.