Make AgentProfile the single user-facing concept. In Settings, collapse Agent / LLM / Condenser / possibly Verification into one Agent configuration page. Users save named AgentProfiles from there. An OpenHands profile can include LLM, condenser, tools/MCP, and verification settings. An ACP profile can include provider, model, command/env, and credentials, but not verification for now. When starting a new conversation, users choose an AgentProfile. That profile fully defines the launch config. Inside an existing conversation, when switching configs, users still choose from AgentProfiles, not raw models. The UI determines which profiles are runtime-compatible with the current conversation. Compatible profiles can be switched live; incompatible ones are greyed out with a reason like “requires a new conversation.” This keeps the UX consistent and avoids silently applying only part of a profile.

One caveat: if you have an openhands profile with specific mcp or tools, then with that design you can only switch to another profile with same tools/mcp but other llm. Because tools/mcp are runtime invariant so the two profiles would appear incompatible.

that problem is solved if tools/mcp are set globally. Then for openhands agent profile is just (agent=openhands, llm) so all openhands agent profiles are compatible