Decouple your AI agent runtime from UI surfaces using immutable event streams
Every agent interaction is captured as an ordered stream of immutable events. Reconstruct state, replay sessions, and audit all behavior.
Agent runtime contains no UI-specific logic. Multiple surfaces — CLI, VS Code, Web, Desktop — project from the same event stream.
Switch between OpenAI, Anthropic, and local models without changing agent code. Provider-agnostic interface with fallback chains.
Abstract interfaces for Git, GitHub, Shell, Kubernetes. Swap implementations, test with mocks, add new capabilities without touching the runtime.
Full session state machine — create, pause, resume, complete, fail, cancel. Sessions are recoverable from event streams after crashes.
Reconstruct any session state from events. Snapshots for fast recovery. Event schema versioning for forward compatibility.
DAG-based causality tracking for agent actions. Classifies messages by type and scores them based on provenance outcomes, reducing token usage by 50-80% while preserving critical context.
Lifecycle hook system for intercepting model calls, tool executions, and session events. Error-isolated hooks with timeout support enable observability and customization without modifying the runtime.
Cost tracking and performance auditing with per-model token metrics. Aggregates model usage and emits attestation events on session completion with full provenance summaries.