Cortavyncortavyn-deep composes Graph, Chat, and the model API into a portable deep-agent harness. It provides a thread-isolated virtual workspace, task planning, todos, persistent memory, discovered skills, static and dynamic subagents, MCP tools and resources, interpreter tools, and approval-aware writes.
invoke and resume run through an internal StateGraph. A pending approval is therefore a durable graph interrupt rather than an in-process callback. stream(DeepRequest) and resumeStream(...) publish typed progress events, including model text deltas when the model supports streaming.
The harness keeps large tool results outside the prompt, applies character- and token-aware context limits, and retries a context-window overflow with a compacted conversation. Static system, memory, and skill context carries provider-specific cache markers by default; disable that behavior with PromptCachePolicy.disabled() when a provider or deployment requires it.
Start with the Deep Agent quickstart.