Agents

Create, run, and manage agents in Msty Claw

Agents are the execution unit in Claw. Each agent combines identity, access, runtime, and behavior so automations stay predictable and easy to operate.

Open Settings > Agents to create and manage them.

What an agent contains

An agent profile usually includes:

  • Workspace
  • Routing mode
  • Runtime mode
  • Permission profile
  • Working style

Think of this as: who the agent is, where it can run, and how it should behave. Clear agent design makes downstream operations easier because behavior becomes predictable across repeated runs.

First reliable agent pattern

  1. Pick workspace scope (managed or project folder)
  2. Choose runtime (Host or Container)
  3. Set routing (Local only or channel preset)
  4. Apply permission and security constraints
  5. Set/confirm working style
  6. Run a safe validation task

This sequence is meant to reduce risk during onboarding. Keep early agents simple so failures are easy to diagnose.

Workspace options

  • Managed: isolated workspace managed by Claw
  • Project folder: execution in existing repo/docs folders

Guidance:

  • Use managed workspaces for isolation and experimentation
  • Use project folders for repo-aware execution
  • Enable Workspace Profile Files only when root-level SOUL.md, IDENTITY.md, or USER.md files should shape behavior in that workspace

When selecting project-folder mode, scope access tightly to the paths required for the job.

Routing options

  • Local only: no remote channel dependency
  • Channel preset: reusable remote destination and sender policy

Start local-only first, then add channel routing once agent behavior is stable. This keeps your first validation focused on execution quality rather than transport and sender-policy complexity.

Runtime options

  • Host: runs directly on your machine
  • Container: runs in Docker/Podman, with OrbStack support on macOS

Container controls include:

  • Network on/off
  • Shell/terminal access
  • Web access behavior (depends on network setting)
  • Optional custom container environments for required tools and packages

Guidance:

  • Use container runtime for stronger isolation
  • Use host runtime for faster local iteration

Pick runtime based on risk profile. If the agent handles sensitive operations, prefer stronger isolation even if iteration speed is slower.

When using container runtime with helper/sub-agent workflows, keep environment configuration explicit so all delegated work runs with the same expected tools.

Computer Use

Claw supports Computer Use on macOS and Windows for compatible agents when enabled. This allows agent workflows to interact more directly with the desktop environment.

Use Computer Use for tasks that genuinely require UI interaction. Keep it disabled for workflows that can be completed through files, APIs, or shell tools.

Operational guidance:

  • Enable Computer Use only on agents with clear ownership
  • Combine with strict sender and permission policies
  • Validate behavior in low-risk scenarios before broader rollout
  • Keep it off for agents that only need shell, file, API, or MCP access

Lifecycle actions

Common operations:

  • Create
  • Edit
  • Clone
  • Open chat
  • Open activity
  • Stop
  • Delete

Treat clone-and-validate as your default change pattern for production agents instead of editing critical profiles in place.

Local chat handoff

You can hand off a local chat to an agent when the conversation has useful context but needs an execution identity.

Handoff can:

  • Attach the existing conversation session to a running agent
  • Preserve the current transcript and session state
  • Carry over provider/model context where available
  • Preserve workspace path and folder access, merging source and target folder access when needed
  • Create a new agent from the chat when no existing agent fits

Use handoff when a local conversation has matured into repeatable or permissioned work. Do not use it for trivial chats where creating an agent would add unnecessary operational state.

Mission Control handoff

Mission Control is where complex multi-agent runs are tracked and reviewed.

  • Start and manage execution identities in Agents
  • Use Mission Control to coordinate fresh-context helpers and approvals
  • Configure reusable helper presets in Settings > Task Force

Operations guidance

  • Clone agents to test behavior without changing production profiles
  • Keep one purpose per agent when possible
  • Review activity after major config changes
  • Re-validate permissions when workspace/runtime changes

As agent count grows, naming and ownership conventions become important. Assign clear purpose and owner per agent to keep operations manageable.

Related detail pages

Security
Configure sender controls, folder policy, and network guard in Msty Claw
Memory Bank
Use Working Brief and Memory Packs for durable context