Agent Config Object

What This Demo Shows

Forwarded props flowing from the React provider into the Microsoft Agent Framework agent's system prompt on every turn. The agent adapts tone, assumed expertise, and response length based on selector state — no code change, no conversation restart.

How to Interact

1. Pick a tone, expertise level, and response length from the config card.
2. Send any message (e.g. "Explain how a database index works").
3. Change one of the selects and send a follow-up ("Explain it again").
4. Compare the two responses — the second reflects the new settings.

Technical Details

Provider properties are passed via <CopilotKit properties={{...}}>. CopilotKit serializes them into the AG-UI forwardedProps field on every agent run.

Backend — src/agents/agent_config_agent.py defines AgentConfigFrameworkAgent, a subclass of AgentFrameworkAgent that overrides run_agent:

1. Reads forwardedProps from the AG-UI input dict with defensive defaults.
2. Composes a system prompt from three small rulebooks (tone / expertise / length).
3. Prepends that prompt as a system message before delegating to the default orchestrator chain.

The underlying MS Agent Framework Agent is static — per-turn customization rides in as an injected leading system message rather than by rebuilding the agent.

Dedicated route — this demo has its own runtime at /api/copilotkit-agent-config proxying to the backend's /agent-config FastAPI endpoint. Scoping keeps the dynamic-prompt behaviour out of the shared agent used by other demos.