LangChain / LangGraph Adapter

Agentomatic ships a library helper — agentomatic.langchain_adapter — that normalises LangChain / LangGraph abstractions (messages, prompt templates, LCEL chains, tools, RunnableConfig) into the platform's dict-based invoke contract.

This is distinct from the Studio LangChainAdapter (agentomatic.studio.adapters.langchain), which powers the debug UI for framework="langchain" agents. Use the library adapter when packaging a LangChain agent as a class agent or bridging message types.

---

Quick start

from agentomatic.langchain_adapter import (
    AgentAdapter, dict_to_messages, messages_to_dict, serialize_messages,
)

# Wrap a compiled LangGraph / LCEL runnable
adapted = AgentAdapter(my_graph_or_chain)
result = await adapted.atransform({"current_query": "Hello", "messages": []})
# Sync entrypoint (safe inside FastAPI / notebooks via run_sync):
result = adapted.transform({"current_query": "Hello"})

# State dict → LangChain messages
msgs = dict_to_messages({"current_query": "Hello", "messages": []})

# Or convert a bare message list (graph dataclass state)
msgs = dict_to_messages([{"role": "user", "content": "Hi"}])

# LangChain messages → state dict / plain list
state = messages_to_dict(msgs)
plain = serialize_messages(msgs)  # list[{"role", "content"}]

Scaffold a LangChain agent

agentomatic init chatbot --template langchain

The template uses dict_to_messages / serialize_messages inside nodes so REST /api/v1/{name}/invoke payloads work with LangChain message objects.

Prompt helpers

from agentomatic.langchain_adapter import (
    resolve_prompt, extract_system_prompt, inject_system_prompt,
)

# Resolve from agent attributes / ChatPromptTemplate / PromptManager
system = resolve_prompt(agent, default="You are helpful.")

# Read or replace the system message on a ChatPromptTemplate / chain
system = extract_system_prompt(prompt_template)
updated = inject_system_prompt(prompt_template, "You are helpful.")

Config helpers

from agentomatic.langchain_adapter import make_config, inject_config

config = make_config(thread_id="t-1", tags=["studio"])
result = await chain.ainvoke(payload, config=config)

# inject_config mutates a *state dict* with runnable_config / thread_id keys
state = inject_config({"current_query": "Hi"}, config)

Adapting LangGraph agents

from agentomatic.langchain_adapter import adapt_langgraph_agent

wrapped = adapt_langgraph_agent(compiled_graph, agent_name="chatbot")

Studio vs library

Surface	Module	Role
Studio debug UI	studio.adapters.langchain.LangChainAdapter	SSE streaming, LCEL graph, traces
Runtime bridge	langchain_adapter.AgentAdapter	Dict ↔ messages, prompt/tools helpers

Both can be used together: wrap with AgentAdapter for REST/optimize, and declare framework="langchain" (for LCEL node_fn agents) so Studio selects its adapter automatically. Class agents (BaseGraphAgent) keep framework="graph_agent" and still use langchain_adapter helpers inside nodes.