Run agents
Start sessions in Orca Agent Engine, stream their output, choose the sandbox an agent executes in, and supply the credentials and events it runs on.
Agents do not run continuously. Every run is a session: a running instance of a pinned agent version inside an environment, the container template that decides what packages are available and what network the agent can reach. Sessions are the unit you create, drive, observe, and tear down.
You drive a session by sending it events and reading the response stream. The same event stream carries tool-call requests, steering messages, and interrupts, so a session is a conversation your application participates in rather than a job you submit and wait on.
Environments
Define the sandbox template - packages and network policy - that sessions execute inside.
Sessions
Create a session, pin an agent version, and mount resources into its sandbox.
Session operations
Retrieve, list, update, interrupt, archive, and delete an existing session.
Multiagent
Delegate to other agents inside one session and read each delegate's thread.
Sandboxes
What the sandbox guarantees: filesystem layout, the writable path, and caps.
Events and streaming
Send events to a session and stream its responses as Server-Sent Events.
Best practices
Credential handling, idempotency, pagination, retries, and environment separation.
Two things a session needs that the agent definition deliberately does not hold. Triggers start sessions on a schedule in every deployment; StreamNative Cloud extends them to Pulsar and Kafka topics, so an agent can react to your systems instead of waiting for an API call. Vaults hold the credentials a session presents to the MCP servers its agent declares, so the agent definition can name a server without carrying a secret.