Skip to content

Agents & LLM

Nine components providing agent interfaces for Claude, OpenAI, LangChain, LangGraph, AutoGen, smolagents, and expert systems — a multi-provider agent layer.

Overview

G6 does not lock into a single LLM provider or agent framework. This cluster provides standardized AIBlock wrappers for multiple agent backends.

agent_claude (Beta) is the most mature implementation: Anthropic SDK, tool calling, autonomous infer_loop() (call → observe → execute → repeat), retry policy, circuit breaker, and lifecycle management. It is the recommended backend for production use.

The remaining wrappers (agent_openai, agent_langchain, agent_langgraph, agent_autogen, agent_smolagents) provide basic single-shot inference through their respective SDKs but do not yet support autonomous tool-use loops. agent_nanoclaw and agent_openclaw are constrained-environment agent/control-plane components (Ollama/OpenRouter only — no Anthropic backend per TOS). NanoClaw's MCP tools currently persist coordination records locally unless a real NanoClaw runtime backend is configured, so task, sandbox, and skill operations should not be treated as completed runtime execution from success=True alone. adapt_experta wraps the Experta rule engine for symbolic reasoning.

Components

Component Description MCP Tools
agent_claude Anthropic Claude agent with tools and streaming --
agent_openai OpenAI agent wrapper --
agent_langchain LangChain agent integration --
agent_langgraph LangGraph stateful agent workflows --
agent_autogen Microsoft AutoGen multi-agent conversations --
agent_smolagents HuggingFace smolagents lightweight agents --
agent_nanoclaw Control-plane coordination ledger for constrained NanoClaw environments; runtime execution requires a configured backend --
agent_openclaw Open-source agent abstraction --
adapt_experta Experta rule engine for expert systems --

Architecture

graph TD
    LLM[llm_router] --> CLAUDE[agent_claude]
    LLM --> OPENAI[agent_openai]
    LLM --> LC[agent_langchain]
    LC --> LG[agent_langgraph]
    LLM --> AG[agent_autogen]
    LLM --> SM[agent_smolagents]
    LLM --> NANO[agent_nanoclaw]
    LLM --> OPEN[agent_openclaw]
    EXP[adapt_experta] --> CORE[core.AIBlock]
    CLAUDE --> TOOLS[Tool Specifications]
    OPENAI --> TOOLS

Key Patterns

Uniform Agent Interface. All agent components accept AgentInput (messages, optional tools, optional config) and return AgentOutput (response text, tool calls, usage info). This lets higher-level orchestrators swap providers without changing calling code.

Mock-Based Testing. Agent tests mock the SDK client rather than making real API calls. For agent_claude, the pattern is patch("mvp.agent_claude.claude_block.anthropic", mock_anthropic). This isolates tests from network dependencies and API key requirements.

Expert System Bridge. adapt_experta provides a different reasoning paradigm -- forward-chaining rule engines rather than LLM inference. This is valuable for domains with well-defined rules (compliance, medical protocols) where deterministic reasoning is preferred over probabilistic generation. For non-developer users, it should be surfaced through templates or natural-language extraction; asking first-time users to hand-author fact schemas and rule objects undermines the under-10-minute onboarding goal.