LangGraph agents

​

Create a LangGraph agent

guild agent init --name my-langgraph-agent --agent-type langgraph

my-agent/
├── langgraph.json    # Names the graph to run
├── graph.py          # The graph
├── guild.yaml        # Integrations, sub-agents, and built-in tools
├── README.md
└── .gitignore

from guildai_langchain import guild_model
from langchain.agents import create_agent

graph = create_agent(
    guild_model(),
    tools=[],
    system_prompt="You are my-agent, an agent running on the Guild platform.",
)

​

The manifest

{
  "dependencies": ["."],
  "graphs": {
    "agent": "./graph.py:graph"
  }
}

• graphs is required. Each target is a relative path to a committed .py file, a colon, and an attribute name, such as ./my_agent/graph.py:graph. Module-style targets (my_agent.graph:graph) are not supported.
• When graphs declares several graphs, Guild runs the one named agent. When it declares one, Guild runs that one.
• dependencies is optional. See Dependencies.

• A compiled graph, such as the one create_agent returns
• An uncompiled StateGraph. Guild compiles it.
• A function that takes no arguments and returns either of the above. It can be async. Use this form when the graph needs Guild tools.

​

Models

​

Tools

integrations:
  - name: acme~github
    version: ^1.4.0
    tools: [github_repos_get, github_issues_list]

sub_agents:
  - name: acme~research
    version: ^1.0.0

builtins:
  - name: ui
    tools: [ui_prompt]

from guildai_langchain import guild_model, guild_tools
from langchain.agents import create_agent

async def graph():
    return create_agent(guild_model(), await guild_tools())

​

Input and output

• Input — each turn’s text arrives as a HumanMessage appended to the graph’s messages. The graph must accept {"messages": [...]}, as create_agent and any MessagesState graph do. A graph with a custom state schema fails its first turn.
• Conversation — Guild saves the message history in the task’s state between turns, so a follow-up message in the same session continues the conversation.
• Output — the text of the model’s last response in the turn. If the model produced no text, the output says so.

​

Dependencies

​

Limits

• Idle turns — a turn fails if the graph produces no events for 10 minutes. The clock pauses while a Guild tool call is in flight, so a long sub-agent call or a wait on ui_prompt doesn’t count.
• Import errors — the build checks the manifest but doesn’t import your code. A syntax error, or a graph that can’t be built, fails the first turn that runs it. Test with guild agent test or guild agent chat before you publish.

​

Build-time validation

1. Validate LangGraph manifest — langgraph.json is present and valid, every graph target points to a committed .py file inside the workspace with a valid attribute name, a graph named agent exists when there are several, and every dependency is available.
2. Parse guild.yaml — guild.yaml is valid YAML, with no unknown keys and only sections a LangGraph agent supports.
3. Validate Guild integrations — each integration resolves to a published version, and any listed tools map to real operations on it.
4. Validate Guild subagents — each sub-agent resolves to a published version.
5. Validate Guild builtins — each service is one of console, ui, or guild, and each listed tool is permitted for that service.
6. Validate LLM models — each entry under models names a supported provider.
7. Store Guild configuration — the resolved tool manifest is recorded on the version.