Features
30-line agent definition (objective, tools, guardrails)
Deterministic guardrails outside the model
Zero-trust agents with RBAC and ABAC
Typed tool contracts and incorruptible domain objects
Multi-agent orchestration with fault recovery
Connectivity: MCP (STDIO + HTTP), A2A, REST, OpenAPI
Legacy protocol support: databases, filesystems, scanned archives
Gated API integration with credential management and quota awareness
Cloud-agnostic deployment (AWS, Azure, GCP, on-prem, air-gapped)
Model-agnostic (any frontier or open model via Bedrock, Vertex)
Immutable audit logs with full reasoning capture
Granular observability across all workflows
Graceful fault recovery and isolated failure domains
Continuous learning from human overrides in production
Unified runtime reduces token cost; adding agents is free
Decorator-based LLM call pattern (@llm.call)
Automatic versioning with @ops.version
Tracing and cost tracking per call via @ops.trace
Built-in tool calling with @llm.tool
Agent loop with tool execution and response resume
Streaming response support
Thinking/reasoning with include_thoughts parameter
OpenTelemetry integration for observability
Structured output via Python type hints
Multi-provider support (OpenAI, Anthropic, Google)
Lightweight, minimal abstraction design
Provider-agnostic unified interface
Simple tool creation with function decorators
Cost tracking per interaction
GitHub open-source (MIT license)
