Comprehensive guides covering every feature of the Orbiter multi-agent framework.
Core Concepts
| Guide | Description |
|---|
| Agents | Creating and configuring agents — the core autonomous unit |
| Tools | Defining tools with the @tool decorator, FunctionTool, or the Tool ABC |
| Running Agents | Executing agents with run(), run.sync(), and run.stream() |
| Streaming | Real-time streaming with TextEvent and ToolCallEvent |
| Structured Output | Constraining agent output to Pydantic models |
Multi-Agent Systems
| Guide | Description |
|---|
| Multi-Agent Swarms | Orchestrating agents with workflow, handoff, and team modes |
| Agent Groups | Parallel and serial execution groups, nested swarms |
Lifecycle and Observability
| Guide | Description |
|---|
| Hooks | Intercepting agent lifecycle events (pre/post LLM call, tool call) |
| Events | Decoupled async event bus for cross-cutting concerns |
Human Oversight
| Guide | Description |
|---|
| Human-in-the-Loop | Pausing agents for human input, confirmation, or review |
Context Engine
| Guide | Description |
|---|
| Context Overview | Architecture, automation modes, and full context engine guide index |
| State Management | Hierarchical key-value state, parent inheritance, fork/merge |
| Prompt Building | Neurons, priority ordering, variable substitution |
| Processors | Event-driven pipeline (pre_llm_call, post_tool_call, etc.) |
| Workspace | Versioned artifact storage, observers, filesystem persistence |
| Knowledge | Text chunking, TF-IDF search, knowledge store |
| Checkpoints | Snapshot save/restore, version history |
| Token Tracking | Per-agent per-step usage, trajectories, summaries |
| Context Tools | Planning, knowledge, and file tools for agents |
Memory
| Guide | Description |
|---|
| Memory | Typed memory hierarchy, short/long-term memory, orchestration, summarization |
| Memory Backends | SQLite, Postgres, and Vector storage backends |
Integration
| Guide | Description |
|---|
| MCP | Model Context Protocol client/server, tool loading, multi-server management |
| Sandbox | Isolated execution environments (local, Kubernetes), filesystem and terminal tools |
| Agent-to-Agent (A2A) | Agent discovery, HTTP-based agent communication, remote agent wrappers |
Observability and Evaluation
| Guide | Description |
|---|
| Tracing | OpenTelemetry spans, @traced decorator, prompt logging, propagation |
| Evaluation | Rule-based and LLM-as-Judge scorers, parallel evaluation, pass@k |
| Ralph Loop | Iterative refinement: Run -> Analyze -> Learn -> Plan -> Halt |
Training
| Guide | Description |
|---|
| Training | Trajectories, data synthesis, evolution, VeRL/RLHF integration |
Deployment
| Guide | Description |
|---|
| Server | FastAPI-based HTTP server with chat, SSE, WebSocket streaming |
| CLI | Command-line runner, interactive console, batch processing, plugins |
Configuration and Extensibility
| Guide | Description |
|---|
| Config-Driven Agents | Loading agents and swarms from YAML files |
| Skills | Skill registry, markdown skill format, local and GitHub sources |
| Model Providers | Using OpenAI, Anthropic, and custom LLM providers |
Quick Reference
All framework types live in orbiter.types:
- Messages:
UserMessage, SystemMessage, AssistantMessage, ToolResult
- Execution:
AgentOutput, RunResult, ActionModel
- Streaming:
TextEvent, ToolCallEvent, StreamEvent
- Stats:
Usage
Configuration types live in orbiter.config:
ModelConfig, AgentConfig, TaskConfig, RunConfigparse_model_string() — split "provider:model" strings
For full API details, see the API Reference.