====== Langroid ======
**Langroid** is an open-source Python framework for building multi-agent LLM applications that treats agents as first-class citizens capable of collaborating through hierarchical task delegation and structured message passing.(([[https://[[github|github]])).com/langroid/langroid|github.com/langroid/langroid]])) Created with a focus on developer ergonomics, Langroid provides a clean, code-driven abstraction where each agent wraps an LLM along with optional tools and vector stores.

  * **[[github|GitHub]]:** [[https://github.com/langroid/langroid|github.com/langroid/langroid]]
  * **License:** MIT
  * **Language:** Python

===== Architecture =====
Langroid's core design centers on two key abstractions:

**Agents:** First-class objects encapsulating LLM conversation state, memory (via vector stores), and tools. Each agent maintains its own context and can be specialized for particular tasks.

**Tasks:** Wrappers around agents that define specific instructions and enable hierarchical, recursive delegation. Tasks can spawn subtasks and delegate work to other agent-task pairs, forming collaborative workflows.

Key architectural features:
  * **Hierarchical delegation:** Agents can recursively delegate subtasks to other agents, forming tree-structured workflows
  * **Pydantic-based tool calling:** [[structured_outputs|Structured outputs]] with automatic validation and error handling
  * **Vector store integration:** Built-in support for [[qdrant|Qdrant]], Chroma, and LanceDB for RAG capabilities
  * **Knowledge graph support:** Integration with [[neo4j|Neo4j]] and ArangoDB for graph-based reasoning
  * **LLM flexibility:** Compatible with any LLM ([[openai|OpenAI]], [[anthropic|Anthropic]], local models, open-source via API)
  * **Redis caching:** Prompt/response caching for performance optimization
  * **Message provenance:** Detailed logging with source tracking for debugging and observability

===== Python Example =====
<code python>
import langroid as lr
from langroid.agent.tools.orchestration import DoneTool

# Configure the LLM (works with [[openai|OpenAI]], [[anthropic|Anthropic]], local models)
llm_config = lr.language_models.OpenAIGPTConfig(
    chat_model="[[openai|openai]]/gpt-4o",
)

# Create a research agent with specific instructions
researcher_config = lr.ChatAgentConfig(
    name="Researcher",
    llm=llm_config,
    system_message="You research topics and provide detailed factual summaries.",
)
researcher = lr.ChatAgent(researcher_config)

# Create a writer agent that refines research into polished text
writer_config = lr.ChatAgentConfig(
    name="Writer",
    llm=llm_config,
    system_message="You take research notes and write clear, engaging prose.",
)
writer = lr.ChatAgent(writer_config)

# Wrap agents in Tasks for orchestration
researcher_task = lr.Task(researcher, interactive=False)
writer_task = lr.Task(writer, interactive=False)

# The writer delegates research subtasks to the researcher
writer_task.add_sub_task(researcher_task)

# Run the pipeline: writer receives a goal and delegates as needed
result = writer_task.run(
    "Write a brief overview of how HNSW graphs work for vector search."
)
print(result)
</code>

===== Use Cases =====
Langroid excels in scenarios requiring fine-grained multi-agent collaboration:

  * **Advanced RAG:** Knowledge bases with cited answers from documents
  * **Business intelligence:** Automated data queries, SQL integration, and workflow automation
  * **Complex problem-solving:** Strategic planning, diagnostics, multi-step reasoning
  * **AI-driven development:** Code generation, analysis, and security auditing
  * **Debate platforms:** Multi-agent setups with dynamic agent parameterization for structured discussions

===== Langroid vs CrewAI vs AutoGen =====
^ Feature ^ Langroid ^ [[crewai|CrewAI]] ^ [[autogen|AutoGen]] ^
| **Interface** | Code-only (no GUI) | Orchestration-focused, integrable | [[conversational_agents|Conversational agents]] |
| **Focus** | Custom multi-agent hierarchies | Production agent fleets | Research-oriented conversations |
| **Customization** | Full Python extensibility | Plugin marketplace, easier setup | Flexible conversation patterns |
| **Best For** | Developers/researchers needing fine control | Teams deploying agent workflows | Conversational multi-agent research |
| **Vendor** | Open-source community | [[crewai|CrewAI]] Inc. | [[microsoft|Microsoft]] |

Choose Langroid when you need full control over agent hierarchies, task delegation logic, and integrations without a GUI layer. Choose [[crewai|CrewAI]] for production orchestration with monitoring, or [[autogen|AutoGen]] for conversational multi-agent patterns.

===== See Also =====
  * [[xagent|XAgent: Autonomous LLM Agent for Complex Tasks]]
  * [[metagpt|MetaGPT: Multi-Agent Framework with SOP-Based Collaboration]]
  * [[langgraph|LangGraph]]
  * [[agentverse|AgentVerse: Facilitating Multi-Agent Collaboration]]
  * [[managed_agents_vs_agent_sdk|Managed Agents vs Agent SDK]]

===== References =====
===== Related Pages =====
  * [[crewai|CrewAI]]
  * [[autogen|AutoGen]]
  * [[multi_agent_systems|Multi-Agent Systems]]
  * [[retrieval_augmented_generation|Retrieval-Augmented Generation]]