CopilotKit

Agent Config

Forward typed configuration from your UI into the agent's reasoning loop.

"""LangGraph agent backing the Agent Config Object demo.Reads three forwarded properties — tone, expertise, responseLength — from theLangGraph run's ``RunnableConfig["configurable"]["properties"]`` dict andbuilds its system prompt dynamically per turn.The CopilotKit provider's ``properties`` prop is wired through the runtime as``forwardedProps`` on each AG-UI run. This graph reads those with defensivedefaults (unknown / missing values fall back to the defaults) and composes thesystem prompt from three small rulebooks before invoking the model."""from typing import Any, Literalfrom langchain_core.runnables import RunnableConfigfrom langchain_openai import ChatOpenAIfrom langgraph.graph import END, START, MessagesState, StateGraph_llm: ChatOpenAI | None = Nonedef _get_llm() -> ChatOpenAI:    """Lazy-instantiate the LLM so importing this module (e.g. in unit tests)    does not require ``OPENAI_API_KEY`` to be set."""    global _llm    if _llm is None:        _llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.4)    return _llmTone = Literal["professional", "casual", "enthusiastic"]Expertise = Literal["beginner", "intermediate", "expert"]ResponseLength = Literal["concise", "detailed"]DEFAULT_TONE: Tone = "professional"DEFAULT_EXPERTISE: Expertise = "intermediate"DEFAULT_RESPONSE_LENGTH: ResponseLength = "concise"VALID_TONES: set[str] = {"professional", "casual", "enthusiastic"}VALID_EXPERTISE: set[str] = {"beginner", "intermediate", "expert"}VALID_RESPONSE_LENGTHS: set[str] = {"concise", "detailed"}def read_properties(config: RunnableConfig | None) -> dict[str, str]:    """Read the forwarded ``properties`` object with defensive defaults.    Any missing or unrecognized value falls back to the corresponding    ``DEFAULT_*`` constant. The function never raises.    """    configurable = (config or {}).get("configurable", {}) or {}    properties = configurable.get("properties", {}) or {}    tone = properties.get("tone", DEFAULT_TONE)    expertise = properties.get("expertise", DEFAULT_EXPERTISE)    response_length = properties.get("responseLength", DEFAULT_RESPONSE_LENGTH)    if tone not in VALID_TONES:        tone = DEFAULT_TONE    if expertise not in VALID_EXPERTISE:        expertise = DEFAULT_EXPERTISE    if response_length not in VALID_RESPONSE_LENGTHS:        response_length = DEFAULT_RESPONSE_LENGTH    return {        "tone": tone,        "expertise": expertise,        "response_length": response_length,    }def build_system_prompt(tone: str, expertise: str, response_length: str) -> str:    """Compose the system prompt from the three axes."""    tone_rules = {        "professional": ("Use neutral, precise language. No emoji. Short sentences."),        "casual": (            "Use friendly, conversational language. Contractions OK. "            "Light humor welcome."        ),        "enthusiastic": (            "Use upbeat, energetic language. Exclamation points OK. Emoji OK."        ),    }    expertise_rules = {        "beginner": "Assume no prior knowledge. Define jargon. Use analogies.",        "intermediate": (            "Assume common terms are understood; explain specialized terms."        ),        "expert": ("Assume technical fluency. Use precise terminology. Skip basics."),    }    length_rules = {        "concise": "Respond in 1-3 sentences.",        "detailed": ("Respond in multiple paragraphs with examples where relevant."),    }    return (        "You are a helpful assistant.\n\n"        f"Tone: {tone_rules[tone]}\n"        f"Expertise level: {expertise_rules[expertise]}\n"        f"Response length: {length_rules[response_length]}"    )def call_model(    state: MessagesState, config: RunnableConfig | None = None) -> dict[str, Any]:    """Single graph node — read forwarded props, build prompt, invoke LLM."""    props = read_properties(config)    system_prompt = build_system_prompt(        props["tone"], props["expertise"], props["response_length"]    )    messages = [{"role": "system", "content": system_prompt}] + state["messages"]    response = _get_llm().invoke(messages)    return {"messages": [response]}graph_builder = StateGraph(MessagesState)graph_builder.add_node("model", call_model)graph_builder.add_edge(START, "model")graph_builder.add_edge("model", END)graph = graph_builder.compile()

You have a working agent and want the user to be able to tune how it behaves: tone, expertise level, response length, language, persona. By the end of this guide, your UI will own a typed config object that the agent reads on every run and rebuilds its system prompt from.

When to use this#

Reach for agent config whenever the agent's behaviour depends on user-controllable settings that don't fit naturally as chat input:

  • Tone, voice, persona: "playful", "formal", "casual"
  • Expertise level: "beginner", "intermediate", "expert"
  • Response shape: short / medium / long, structured / prose, language
  • Domain switches: which knowledge base to consult, which tool subset to enable

If the values are a channel the user occasionally tunes (a settings panel, a toolbar of selects), agent config is the right shape. If the values are content the agent should write back to (notes, a document, a plan), use Shared State instead.

How agent config flows from the UI into the agent's reasoning loop depends on your runtime architecture. Agents living behind a runtime read it from agent state on every run, while in-process agents receive the same object as forwarded properties on the provider — same UX, slightly different wiring on each side.

How it works#

Install the LangGraph Python SDK

uv add copilotkit
poetry add copilotkit
pip install copilotkit --extra-index-url https://copilotkit.gateway.scarf.sh/simple/
conda install copilotkit -c copilotkit-channel

Wire CopilotKit middleware into your graph

Agent config flows from the UI into the agent via useAgentContext — the frontend publishes a typed object and CopilotKitMiddleware injects it into the model's prompt on every turn. Make sure the middleware is in your create_agent call.

agent_config_agent.py
graph = create_agent(
    model=ChatOpenAI(model="gpt-5.4", temperature=0.4),
    tools=[],
    middleware=[CopilotKitMiddleware()],
    system_prompt=SYSTEM_PROMPT,
)

Read the resulting config inside your system prompt or a custom middleware — see src/agents/agent_config_agent.py for the full rulebook-driven shape used in the showcase.

Agent config is a typed object the frontend owns and publishes to the agent as runtime context. There are two pieces: the UI side, which owns the React state and publishes every change with useAgentContext, and the backend node, which reads that context entry and turns it into a system prompt.

The UI side stays simple. Hold the typed config in React state, then mirror every change into the agent through useAgentContext:

frontend/src/app/page.tsx — UI publishes the typed config
function ConfigContextRelay({ config }: { config: AgentConfig }) {
  useAgentContext({
    description: "Agent response preferences",
    value: {
      tone: config.tone,
      expertise: config.expertise,
      responseLength: config.responseLength,
    },
  });
  return null;
}

The backend half is also a single node. Read the latest config context at the top of every run and use it to build the system prompt for that turn:

backend/agent.py — agent reads config and rebuilds the system prompt
import json

CONFIG_KEYS = ("tone", "expertise", "responseLength")

def read_config_value(entry):
    value = entry.get("value")
    if isinstance(value, str):
        try:
            value = json.loads(value)
        except json.JSONDecodeError:
            return None
    if not isinstance(value, dict):
        return None
    if any(key in value for key in CONFIG_KEYS):
        return value
    return None

async def my_agent_node(state: AgentState, config: RunnableConfig):
    context_entries = state.get("copilotkit", {}).get("context", [])
    cfg = next(
        (
            value
            for entry in reversed(context_entries)
            if (value := read_config_value(entry)) is not None
        ),
        {},
    )
    tone = cfg.get("tone", "professional")
    expertise = cfg.get("expertise", "intermediate")
    response_length = cfg.get("responseLength", "concise")
    system_prompt = build_system_prompt(tone, expertise, response_length)
    # ...

The agent reads the latest typed config at the start of every turn, rebuilds the system prompt, runs the turn. This is the same shape as the shared-state write-side pattern; agent config is just a specific use of that pattern with a UI-owned typed object on top.

On this page

When to use thisHow it works