Human-in-the-Loop (HITL) Traits

Human-in-the-Loop traits connect autonomous agent workflows to real humans — capturing feedback, routing decisions for human review, and incorporating biometric signals into scene logic. These traits bridge the gap between fully autonomous AI and human oversight.

---

Trait Reference

@hitl

Marks an object or action as requiring human confirmation before proceeding. The agent pauses, presents a review request to a designated human reviewer, and resumes only after approval.

object "DeployGate" @hitl {
  geometry: "cube"
  color: "orange"

  hitl: {
    reviewer: "ops-team@example.com"
    timeout_ms: 300_000       // 5 minutes — auto-reject on timeout
    message: "Review before deploying to production"
    urgency: "high"
    escalation: "cto@example.com"
  }

  onGrab: {
    // Pauses here until human approves
    const decision = await hitl.requestApproval({
      context: current_deployment,
      risk_score: agent.risk_assessment()
    })

    if (decision.approved) {
      deploy.toProduction()
    }
  }
}

@feedback_loop

Captures structured human feedback on agent outputs and feeds it back into the agent's learning system. Supports thumbs-up/down, ratings, and free-form corrections.

object "AIResponseCard" @feedback_loop {
  geometry: "plane"
  scale: [1, 0.5, 1]

  feedback_loop: {
    modes: ["thumbs", "rating", "correction"]
    agent_id: "scene-generator-v2"
    store_to: "supabase"         // "local" | "supabase" | "s3"
    use_for_finetuning: true
  }

  onFeedback(feedback) {
    agent.recordFeedback(feedback)

    if (feedback.type === "correction") {
      agent.applyCorrection(feedback.correction)
      scene.rebuild()
    }
  }
}

@biofeedback

Reads physiological signals from connected wearables (heart rate, EEG, GSR, eye tracking) and uses them as scene inputs. Adapts the experience in real-time based on user state.

composition "AdaptiveTraining" {
  object "StressMonitor" @biofeedback {
    biofeedback: {
      sources: ["heart_rate", "gsr", "eeg_alpha"]
      sampling_rate: 10         // Hz
      providers: ["polar_h10", "muse_2", "shimmer"]
    }

    onBiofeedbackUpdate(reading) {
      if (reading.stress_level > 0.8) {
        scene.setIntensity("calm")
        audio.play("breathing_guide.mp3")
      } else if (reading.stress_level < 0.2) {
        scene.setIntensity("challenge")
      }
    }
  }

  object "FlowDetector" @biofeedback {
    biofeedback: {
      sources: ["eeg_theta", "eeg_alpha"]
    }

    onBiofeedbackUpdate(reading) {
      if (reading.flow_state > 0.7) {
        // User is in flow — increase complexity
        scene.progression.advance()
      }
    }
  }
}

---

HITL Workflow Pattern

Agent generates output
         │
         ▼
  @hitl pauses execution
         │
         ▼
  Human receives review request (email / Slack / in-scene UI)
         │
    ┌────┴────────────┐
    ▼                 ▼
 Approved          Rejected
    │                 │
    ▼                 ▼
Agent continues   Agent logs rejection,
                  requests rework

---

Biofeedback Devices

Device	Signals	Transport
Polar H10	Heart rate, HRV	Bluetooth LE
Muse 2	EEG (alpha, theta, beta)	Bluetooth LE
Shimmer3 GSR+	Galvanic skin response, temperature	Bluetooth/USB
Tobii Eye Tracker	Gaze, pupil dilation	USB
Meta Quest (built-in)	Eye tracking, face tracking	OpenXR ET ext

---

Use Cases

Domain	Trait	How
Medical training	@biofeedback	Adapt simulation difficulty to trainee stress/engagement
AI safety	@hitl	Human confirmation before AI deploys changes
Research	@feedback_loop	Capture expert corrections to fine-tune models
Accessibility	@biofeedback	Detect fatigue and reduce cognitive load automatically
Enterprise AI	@hitl	Compliance gate for AI-generated decisions

---

Related

• AI Behavior Traits
• AI Autonomous Traits
• Agent Framework
• uAA2++ Protocol