The resource model

One run. It pins an agent version, runs in an environment, and keeps an append-only log of events. Many sessions can run the same agent.

The definition: a model, a system prompt, tools, MCP servers and skills, under a name. Every update makes a new version.

The template a session’s sandbox is built from. Every session names one; it is shared across sessions and read again at each turn.

What a session’s log is made of: the messages your application sends, and everything the agent and the session report back.

Starts sessions on a cron schedule, each against the agent version the trigger pinned.

Holds credentials for the MCP servers an agent calls. A session references vaults by id and is never handed a secret.

An uploaded file that a session mounts into its sandbox, read-only to the agent.

A directory of notes that persists across sessions. A session mounts it, and the agent reads and writes it like any other folder.

A bundle of instructions and files an agent opens when a task calls for it. Listed on the agent; pinned to a version when a session starts.

• 1 Create the agent

  A name, a model and a system prompt, plus tools, MCP servers and skills if it needs them. The Registry stores it as version 1.

• 2 Change the prompt or the model
• 3 Change the harness

• 1 Environment

  environment_id, required. The sandbox is built from it when the session first has work to do.

• 2 Vault
• 3 File
• 4 MemoryStore
• 5 Skill

• 1 From your application

  user.message starts or redirects a turn, user.tool_confirmation answers a request for approval, and user.custom_tool_result returns a custom tool’s output.

• 2 From the agent
• 3 From the session itself
• 4 Timing and usage

• 1 A trigger is a schedule

  It names an agent, a cron expression, a message to send, and the environment and vaults its sessions should use.

• 2 Each firing is a new session
• 3 On a pinned agent version