Skip to Content
AgentsCreating Agents

Creating Agents

You create agents by describing the work in Chat — not by filling in forms. Tulsk drafts the agent; you review one card and click Create.


From One Sentence to a Working Agent

Describe the outcome

In Chat, type what you want handled:

“Answer customer support emails and escalate refund requests to me.”

Review the agent card

Chat drafts a complete agent and shows an editable card: name, role, what it does, the engine it runs on, and the model. Open “Edit instructions” to tune its system prompt. Everything has sensible defaults.

Click Create

The agent is created and shows up under Agents in the sidebar as a persistent teammate. If a connection is needed (an inbox, a calendar), Chat asks for that one thing inline.

Building a team? Describe the whole thing (“set me up a support agent, a researcher, and a PM”) — Chat proposes the team and stacks one card per agent so you confirm each.


Choosing an Engine

The card’s engine picker only shows what your workspace has enabled:

EngineWhat it is
OpenClawA dedicated container — browsing, files, full toolset. The default for heavy lifting.
Direct APILightweight — calls the model API directly. Fastest to start.
Claude ManagedRuns on Anthropic-managed infrastructure.

You can also pick the model (GPT, Claude, GLM, Kimi…) per agent. Don’t want to choose? The defaults are good — that’s the point.


After Creation

Your agent appears on the Agents page with quick actions: Chat · Configure · Issues · Triggers · Channels (power settings live behind Advanced). You can:

  • Talk to it from the main chat with @AgentName — see Chat
  • Give it triggers so it runs on a schedule
  • Teach it skills from the registry
  • @-mention it on any task comment to put it to work on that task

Paid plans include up to 6 agents. The first agent on a paid plan provisions its runtime automatically after checkout.

Last updated on