"Unified Agentic Control Surface Research"

Unified Agentic Control Surface Research (April 2026)

Overview

This research document synthesizes industry standards for Human-in-the-Loop (HITL) steering, the "Reflection Pattern" (Self-Reflection and Verification), and how these concepts map to and unify Vox's existing ecosystem constraints. The goal is to provide a single, unified mental model for the "Pilot Console"—the primary interface through which a human orchestrates the AI system.

This document builds upon previous research, specifically the L.A. Noire Doubt Metaphor and Continuation Prompt Engineering.

Core Concepts & Industry Alignment

The "Reflection Pattern" (Generate-Validate-Reflect)

Modern autonomous coding agents (e.g., LangGraph, smolagents, OpenHands) rely heavily on a cyclical reasoning process:

  1. First Pass (Generate): The agent generates an initial attempt based on the intent (starter prompt).
  2. Validator (Test): An automated execution environment or linter runs against the generated output to gather ground truth.
  3. Second Pass (Reflect): The agent ingests the error logs or validation failures, acting as a debugger to refine its initial attempt.

The "Second Pass" is where reliability jumps from simple text prediction to robust software engineering.

Human-in-the-Loop (HITL) Steering

Effective HITL shifts control from micro-management to delegation and oversight. The control surface must allow humans to define goals, monitor progress, inject suspicion, and halt the system.

Unifying Vox's Control Surface: The Tri-State Pilot Console

We must distill Vox's various control vectors (Starter Prompts, Planning Prompts, Continuation Prompts, Suspicious/Doubt signals, validation rules, and Stop commands) into the smallest possible cognitive footprint for the operator.

We propose the Tri-State Pilot Console:

State 1: Strategic Thrust (Launch & Steer)

This is the system's forward momentum. The human defines what to do and keeps the agent moving.

  • Concepts Unified: Starter Prompt, Planning Prompts, Continuation Prompts.
  • Behavior: The agent is operating in "Generation" mode (First Pass). The UI focuses on delegation.
  • Implementation: The Continuation Prompt acts as the engine oil here, injected periodically to prevent context rot and enforce parallel bulk actions.

State 2: Reflective Interrogation (Doubt & Audit)

This state resolves the conflict between the L.A. Noire "Doubt" metaphor and the "Second Pass Verification." They are the same action.

  • Concepts Unified: L.A. Noire "Suspicious" / "Doubt", Second Pass Validator, Socrates Output-Evaluation.
  • Behavior: When the operator presses "Doubt" (or the system self-triggers doubt due to low Socrates scores), the orchestrator pivots rather than halting. It shifts from generation to Reflective Validation.
  • The Action: The agent explicitly queries the codebase to verify its own recent diffs, runs tests, and applies hallucination checks.
  • UI Representation: Amber heartbeat/pulse. The human says, "I don't trust this," and the machine does the hard work of proving it.

State 3: Circuit Breakers (Halt)

Immediate, non-negotiable stoppage.

  • Concepts Unified: Stop command, Budget Exhaustion, Catastrophic Regression.
  • Behavior: Execution halts entirely. The human must intervene to unblock the loop.
  • Implementation: Red friction UI. Halts the orchestrator's event loop.

Design Decisions: Unifying "Doubt" and "Second Pass"

Historically, Vox treated "Suspicious" (a vague human feeling) and "Improve/Audit" (a concrete action) as separate. Industry research strongly suggests they should be linked.

If the human interface provides a "Doubt" button, it should automatically trigger the "Second Pass" reflection loop. The system should switch models (e.g., to a high-reasoning tier), ingest its own output, and execute the local test verification vox ci check.

By unifying these, we minimize the UI options for the controller while maximizing the automated response to human intuition.

Actionable Guidelines

  1. Reduce Buttons: The UI should primarily feature elements that map cleanly to Start/Continue, Doubt (Verify), and Stop.
  2. Expose Confidence (Socrates): To guide the manual "Doubt" action, the UI should surface the latent Socrates heuristic score so the operator knows when to be suspicious before bugs compound.

References