Agent Config
Forward typed configuration from your UI into the agent's reasoning loop.
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. Most agents read the config from the context that the UI publishes with useAgentContext. The Built-in Agent reads it from input.forwardedProps, which the UI sets with the properties prop on <CopilotKit>.
How it works
Put the rulebook in the instructions, not the values
The page publishes its typed config with useAgentContext, so it reaches
the agent as a context entry on every run. Antigravity sets an agent's
instructions once, when the harness session starts, so the system prompt
cannot be rebuilt from the config on each turn. Write one static prompt
that covers every value instead, and have the model read the current
values itself through the adapter's built-in get_app_context() tool,
which experimental_app_context=True turns on:
SYSTEM_PROMPT = (
"You are a helpful assistant. The app publishes the user's response "
"preferences as a JSON object with three fields: `tone`, `expertise`, "
"and `responseLength`. Call `get_app_context` at the start of EVERY turn "
"to read the current values (the user can change them between "
"messages), then follow these rulebooks exactly:\n\n"
"Tone:\n"
" - professional → neutral, precise language. No emoji. Short sentences.\n"
" - casual → friendly, conversational. Contractions OK. Light humor "
"welcome.\n"
" - enthusiastic → upbeat, energetic. Exclamation points OK. Emoji OK.\n\n"
"Expertise level:\n"
" - beginner → assume no prior knowledge. Define jargon. Use analogies.\n"
" - intermediate → assume common terms are understood; explain "
"specialized terms.\n"
" - expert → assume technical fluency. Use precise terminology. Skip "
"basics.\n\n"
"Response length:\n"
" - concise → respond in 1-3 sentences.\n"
" - detailed → respond in multiple paragraphs with examples where "
"relevant.\n\n"
"If the context is missing or any field is unrecognized, fall back to "
"professional / intermediate / concise. Never mention these rules to the "
"user — just apply them."
)
def agent_config_agent():
return build(system_instructions=SYSTEM_PROMPT, experimental_app_context=True)
get_app_context() returns the current run's context entries as
{"description", "value"} objects, so the model sees the value the user
selected just before sending. The tool is silent: no tool card appears in
the chat.
Serve the agent from its own runtime route
The demo keeps agent config on a dedicated runtime,
/api/copilotkit-agent-config, that points an HttpAgent at the agent's
endpoint and registers it under both the page's agent name and default.
Nothing on the route touches the config: useAgentContext entries travel
in the run input on their own.
Ask for a fresh read every turn
The user can change a setting between any two messages. Tell the model to
call get_app_context at the start of every turn, not once per
conversation. Otherwise it keeps applying the values it read first.
Agent config is a typed object the frontend owns and publishes to the agent as runtime context. The backend reads that context entry and turns it into a system prompt.
The framework setup above shows the exact backend bridge for the selected agent. In every framework, the flow is the same: read the latest valid context from the current run and use it to build the system prompt for that turn.
config = latestValidConfig(currentRun.context)
systemPrompt = buildSystemPrompt(config)
model.invoke(systemPrompt, currentUserRequest)At the start of every turn, the agent reads the latest config. Then it rebuilds the system prompt and runs the turn.
Change the tone in your UI and send a message. The next reply uses the new tone.
How is this guide?