Skip to main content
Three commands control how the agent executes, not what it does. Use them to switch between solo and swarm, force isolated runs, and require auditable proof.

​
/mode — solo vs swarm

By default an agent runs in solo mode — it handles everything itself. Swarm mode lets the agent delegate work to other deployed agents, enabling multi-agent collaboration on complex tasks.
In swarm mode the stream includes child-agent events (swarm_child_start, swarm_child_delta, swarm_child_done) for each delegated task. See Stream Protocol for details.

​
/sandbox — isolated execution

The sandbox command forces the agent to run tasks in an isolated environment. This is useful when the agent processes untrusted input or runs code that should not touch the host.
When sandbox is on, every task runs in isolation. If a sandbox is unavailable, the agent reports an error rather than silently falling back to in-process execution.

​
/proof — auditable execution

Proof mode makes the agent collect a verifiable bundle for each run — input and output hashes per step, inference run IDs, and optional sandbox metadata. Turn it on when you need an audit trail.
Proof bundles are hash artifacts. They let you verify that step inputs and outputs match what the agent claims, but they are not trusted-execution-environment attestations.

​
History commands

Three commands list past activity for the current thread. Each returns a plain text response in the stream — no action events.

​
/thread — past conversations

Lists previous conversation threads associated with this agent.

​
/artifacts — generated media

Lists media the agent has generated — images, audio, video — along with references to retrieve them.

​
/receipt — payment receipts

Lists x402 payment receipts for the thread, including amounts and settlement details. See Payments for the receipt data model.

​
Combining settings

You can stack these commands. A common secure configuration looks like: