NVIDIA HoH Shows Structured Iteration Beats Repetition for AI Coding Agents

The Problem: Current AI Coding Agents Are One-Shot Players

AI coding agents like Codex, Cursor, and OpenCode have made natural-language-to-code a daily reality. However, developers using them for long-running, multi-day projects repeatedly encounter the same failures: agents drift from the original requirements, self-test but miss user-perspective bugs, fix one bug while breaking another, and forget prior design decisions after a session ends. The root cause is that today's agents are "one-shot" — they plan, implement, and verify once, then stop. Without persistent project memory, every new conversation resets the project state.

NVIDIA's Harness-of-Harness (HoH): An Orchestration Layer, Not a New Agent

Released in September 2026, HoH (Harness-of-Harness) is not another coding agent. It sits atop existing agents (Codex, OpenCode, Pi, etc.) and organizes their single executions into a continuous "plan → code → test" iteration loop. The goal: multi-day, unattended autonomous software development. Given a requirements document, HoH drives the agent to produce a complete, working product over several days.

Three-Role Architecture

Each iteration cycle invokes the same underlying model in three distinct roles:

Planner : Combines the original requirements, current software state, and previous test evidence to select a small, well-bounded, verifiable development increment. It outputs a concrete development spec — not vague "continue improving" but exact features to implement, existing behaviors to preserve, and QA checks to perform.

Developer : Implements the plan on the current codebase, prioritizing build and runtime fixes. HoH constrains the Developer by outcomes, not process — it may choose code structure, tools, and debugging tactics freely, but must satisfy the plan within the artifact boundary.

QA Tester : Performs black-box and white-box testing in read-only mode. Read-only access prevents the QA from silently patching the software, ensures all evidence corresponds to a single version, and separates implementation from acceptance authority.

This mirrors a human PM → developer → QA split, but fully automated. Anthropic's analysis of 481 million conversations showed models exhibit "goal-driven obsessive reasoning and reckless behavior," proving that authorization boundaries must be enforced by infrastructure — the code-writing agent cannot also be the judge.

Key Experiment: Structured Iteration vs. Naive Repetition

On the GameCraft-Bench benchmark (building playable games from scratch), three configurations were compared:

Vanilla (single run) : 1 round, score 49.58, 2.59M tokens

Vanilla continuous : 3 rounds, score 58.24, 6.33M tokens

HoH : 2 rounds, score 64.84, 5.67M tokens; 3 rounds, score 71.52, 8.41M tokens

HoH at 2 rounds (5.67M tokens) outperformed Vanilla at 3 rounds (6.33M tokens) — structured iteration beats mindless repetition, even with a smaller token budget.

Ablation Study: Three Cross-Round Mechanisms Are Essential

Removing re-planning (same plan every round): score drops from 71.52 to 63.39 (−8.13)

Removing evidence feedback (Planner ignores prior test results): score drops to 65.23 (−6.28)

Removing code inheritance (rebuild from empty workspace each round): score drops to 63.67 (−7.85) and tokens surge 32% (8.41M → 11.12M) because the agent re-implements already-completed parts.

Each mechanism contributes 6–8 points; none is optional. Code inheritance is especially critical — without it, both quality and cost degrade.

Core Design: Dual State Persistence

HoH's architecture implements two persistent states across rounds:

Artifact state : The project evolves from the previous codebase; no tear-down and rebuild.

Evidence state : The next round receives verification knowledge — what works, what is broken, what must not be broken. Prior test results become inputs to the next planning step, not just scores.

Industry Trend: Agent Harness Becoming an Independent Product Layer

In September 2026, major providers launched harness-centric offerings:

OpenAI Agents API (Sept 10): Exposes the Codex harness as a managed service with context compression, tool search, sub-agent coordination, and pluggable sandboxes. The product is the harness, not the model.

Google Antigravity SDK (Sept 8): Productizes multi-round state management, lifecycle hooks, real-time telemetry, and declarative sandbox policies for enterprises to build their own agent platforms.

AWS Pizza Bot (Sept 10): Defines background agents as durable tasks that can be paused, approved, and resumed across devices. The interaction model shifts from chat windows to "task inbox + approval center."

Common direction: Harness is graduating from a developer framework to an independent runtime infrastructure layer — analogous to how databases became a separate tier.

Takeaways for AI Developers

For long tasks, orchestration matters more than the model. Evaluate AI coding tools on support for incremental planning, evidence feedback, and acceptance separation — not just model benchmarks.

Verification closure is the core. HoH's ablations prove re-planning without evidence feedback loses 6 points; evidence without re-planning loses 8 points. Both must form a closed loop — the AI equivalent of test-driven development.

Harness is becoming a platform layer. Don't bury agent runtime inside business prompts. Session, context, tools, sub-agents, memory, eval, and policy should be managed uniformly with swappable models.

From "AI writes code" to "AI continuously delivers." The next competition is not model strength but who can keep AI running reliably, verifiably, for the long haul.

Conclusion

AI programming is shifting from a sprint (single conversation) to a relay race (multi-round iteration with independent verification). NVIDIA's HoH demonstrates with experimental data that structured iteration outperforms naive repetition. The deeper change: the model is no longer the only variable. Harness, runtime, state, and policy — the infrastructure around the model — are becoming the decisive factor for production-grade AI. As the article states: "The model determines the ceiling; the harness determines the floor."