What the runtime gives you
- Durable execution and context management. Threads recover from transient errors — bad LLM responses, timed-out tool calls, dropped connections — without losing state. Old turns are compacted and sub-agents spun up so threads stay responsive as they grow.
- A cloud sandbox. Each agent has a CLI environment to write and run code in.
- Secure credentials. Connections (OAuth or API keys) are stored encrypted; tools call through them, so the model never sees raw secrets.
- Approvals. Tools can be marked approval-required. The runtime pauses on those calls until a decision is recorded.
- Memory. Agents can write and search structured memory across threads.
What you provide
- Custom tools — HTTP endpoints or MCP servers the agent can call.
- Goals and instructions — system prompt, default model, and standing context the agent can rely on.
- Approval policies — which tools run autonomously and which need a human.
- Inbound events — webhooks or schedules that wake an agent up.
Core objects
- Agent — a durable workspace, usually one per end-user. Holds configuration, memory, connections, and any threads it has run.
- Thread — one conversation or run inside an agent. Threads stream events, can pause for approvals, and can spawn sub-threads.
- Fleet — shared configuration for a group of agents: tools, OAuth apps, default instructions, approval defaults. Configure tools once at the fleet level instead of per agent.
Next steps
- Quickstart — verify your API key, create an agent, start a thread.
- Client libraries — install the TypeScript, Python, Go, Ruby, or CLI client.
- Authentication — API keys, connections, and credential scoping.
- Agents, threads, tools, approvals, and webhooks for the mental model.
- API Reference — exact request shapes and response schemas.