cabal/docs
How Cabal agents work — identity, configuration, and how they fit together.

Agents

A Cabal agent is a persistent identity that owns its own system prompt, control profile, guardrails, wallet bindings, and operating history. When you talk to Cabal in Chat, you are talking to a specific agent. When background work runs, it runs under a specific agent.

One account can own multiple agents. Each agent operates independently — its own conversation history, positions, activity log, and notification settings.

What an agent owns

ComponentWhat it does
System promptSaved trading preferences that shape every response and background run
Control profileHow much authority the agent has: Suggest Only, Guarded, or Full Access
GuardrailsHard and soft limits on execution — position sizes, venues, slippage
Wallet bindingsWhich wallets the agent can trade from, per chain
ExtensionsCabal-curated capabilities enabled on the Bindings tab
Model defaultWhich model the agent uses for new work

Creating an agent

Open the dashboard and go to the agent hire flow. You pick a strategy template (or start from scratch), choose a handle, write a system prompt, and configure readiness checks — provider, wallet, and guardrails.

After deployment, the agent's page has four tabs:

  • Overview — performance, recent activity, status
  • Control — control profile and guardrail settings
  • Instructions — system prompt editor
  • Bindings — provider, wallet, and extension access

Agent selection

The agent selector in Chat and the dashboard determines which agent is active. Switching agents changes:

  • The conversation history you see
  • The system prompt and control profile in use
  • The positions and activity log
  • The wallet bindings and guardrail settings

Omitting an agent selection uses your account's default agent.

How agents execute

Agents do not run as always-on processes. They execute in response to triggers:

  • Interactive — you send a message in Chat
  • Background — a routine fires, a source delivers a signal, a watch condition is met, or a group message arrives

Each trigger starts a run. The run follows a preparation → reasoning → action → completion cycle. See for the full lifecycle.

Control profiles

Your control profile determines what happens when the agent wants to act:

ProfileBehavior
Suggest OnlyAgent prepares proposals. You approve or reject each one.
GuardedAgent executes within guardrail limits. Violations require approval.
Full AccessAgent executes broadly. Guardrails still apply but with fewer restrictions.

The control profile is set per agent, not per conversation. Change it on the agent's Control tab.

FAQ

Can I have different agents for different strategies?

Yes. Create separate agents with their own system prompts and guardrails. One might run a conservative LST strategy while another watches momentum plays.

Do agents share wallets?

Agents can bind to the same wallet, but each agent's positions and activity are tracked separately. The guardrails on each agent apply independently.

What happens when I switch agents in Chat?

The conversation history, system prompt, positions, and activity all switch to the selected agent. Your previous agent's state is preserved — you can switch back.

See also

  • — how runs work step by step
  • — writing effective agent instructions
  • — hard and soft execution limits