Human in the loop

A tool can ask before it acts. loopl pauses that one call inside the agent, emits an event and resumes when a person answers. Other tools in the same turn keep running.

Ask

A Tool returns an ApprovalRequest from requiresApproval(for:). HttpTool does for every request, graded by method — GET low, POST · PUT · PATCH medium, DELETE high — with credential headers redacted. open_app asks unless the scheme is allow-listed.

ApprovalPolicy decides what actually reaches the person: .mutating (the app's default) lets reads through and asks for writes, .always asks for everything, .never never asks. Trusted hosts pass.

Answer

import Loopl

func run(_ model: any Model) async throws {
    let agent = Agent(model: model, tools: [HttpTool()],
                      hooks: [ApprovalPolicy(mode: .mutating, trustedHosts: ["api.example.com"])],
                      approvalTimeout: 120)
    for try await event in agent.stream("Post today's summary to the notes API.") {
        switch event {
        case .approvalRequested(let request):
            await agent.resolve(approval: request.id, .approve)   // or .deny(reason:), .approveEdited(input)
        case .toolResult(let result) where result.isError:
            print(result.content)                                  // a denial reaches the model as an error
        default: break
        }
    }
}

No answer within approvalTimeout (120 s) denies; so does cancelling. agent.pendingApprovals lists what is waiting. An approval is an interrupt underneath.

In SwiftUI

AgentTranscriptView renders an ApprovalCard inline — method, host and path, redacted headers, the body, and Approve · Deny · Edit. Route answers with .onApproval { request, decision in … }. It springs in with LooplDesign.Motion and cross-fades under Reduce Motion.