THE HUMAN–AGENT EXPERIENCE LAYER
Designing steerable interfaces, deterministic boundaries and trust protocols for enterprise AI adoption.
Human–Agent Interaction
2026
Case Study
(00)
THE PROBLEM
The blocker is steerability, not model capability
Enterprise AI adoption stalls less often on model capability than on an interface and steerability crisis. When teams integrate LLMs and autonomous agents into production workflows, three failure modes recur.
Opaque execution
Non-technical domain experts — designers, PMs, ops leads — lack real-time visibility into agent reasoning, which breeds distrust.
Brittle context
Unconstrained agents generate code or content that breaks architectural patterns and violates brand design tokens.
All-or-nothing autonomy
Systems force a false binary between manual labor and unmonitored autonomous execution.
(01)
RESEARCH SCOPE
Twelve months,
three axes
Over 12 months across production client deployments at VML Middle East, enterprise MarOps projects at Hovi Digital Lab / GrowerAI, and self-directed R&D.
The investigation covered human-agent interaction patterns across three axes: steerability protocols (how operators guide agent state without manual code rewriting), deterministic verification (combining LLM output with static AST/schema validation), and visual inspection (real-time feedback for high-dimensional model output).
(02)
THREE FINDINGS
Three findings
Interactive checkpoints outperform post-hoc audits
Human-in-the-loop steering measured 4x more effective when applied at decision checkpoints during agent execution than when reviewing completed output after the fact.
Schema-constrained protocols eliminate the majority of regressions
Enforcing explicit, machine-readable UI contracts at the tool boundary — the approach shipped as cox-protocol — eliminated 80% of the regressions caused by agent context drift and design-token violations.
Structured visual inspection builds operational trust
Live diff previews and activation maps (the approach behind nebulai’s interpretability visualizer) turn agent behavior from an unpredictable black box into something a non-technical reviewer can actually inspect.
(03)
REFERENCE
IMPLEMENTATIONS
Three shipped artifacts carry the argument — each one built, not proposed
cox-protocol
Architected the npm CLI and JSON Schema validator for agentic UI generation. Built with TypeScript, Node.js, JSON Schema.
84% adoption across 35+ engineers and designers at VML Middle East; 40% reduction in visual regression bugs.
nebulai
Built the real-time WebGPU visualization engine for token / SAE feature activations. Built with Python, WebGPU/WGSL, React, UMAP.
Renders 50,000+ active nodes at 60 FPS for interactive interpretability.
career-ops
Implemented the multi-agent pipeline engine for document parsing and match evaluation. Built with Go, JavaScript, Claude Code CLI.
10x speedup in automated workflow execution and template rendering.
(04)
WHY IT MATTERS
Months to days,
and a trail to show for it
Guardrailed interaction patterns cut enterprise AI onboarding from months to days. Deterministic schema enforcement ahead of the API call removes the repetitive agent retry loops that burn tokens without producing usable output. And an immutable trail of human approvals and steering decisions is what emerging AI-safety disclosure requirements — the EU AI Act among them — actually ask for.
The Grower AI
AI Product