Build the orchestrator

@MainActor
final class AgentOrchestrator {
    let state = AgentState()
    var log: [String] = []
    let maxSteps = 6

    func run(userText: String) async throws -> String {
        state.mode = .gather
        var last = userText
        var steps = 0
        while state.mode != .done, steps < maxSteps {
            steps += 1
            let session = LanguageModelSession(profile: AgentProfile(state: state))
            let response = try await session.respond(to: last)
            last = String(describing: response)
            log.append("step \(steps) mode=\(state.mode.rawValue)")
        }
        if state.mode != .done {
            log.append("stopped at cap")
        }
        return last
    }
}

Creating a new session each step is simple and honest. If the SDK keeps transcript across Profile changes on one session, prefer that and pass the same session in. Confirm which pattern the installed docs use. Do not copy a LangGraph checkpointer.

Pack state-specific facts before respond using the context packer. Gather gets search hits. Act gets the chosen action only.

Next, make the log something support can read.