Stop buying smarter models. Start building better AI employees.
NVIDIA’s AVO research is not a new chatbot. It is experimental agent architecture, and it just showed that the harness, not the model, is what turns intelligence into finished work.
Dated Aug 21–22, 2026 · public set only · NVIDIA says this is not a controlled ablation
Pretty demos are easy. A complete agent is the hard part.
Most buyers are shopping for the smartest model. NVIDIA just published the counter-argument: Claude Opus 5 by itself sat at about 30% on ARC-AGI-3. Wrapped in AVO, short for Agentic Variation Operators, the same model family completed the entire public set.
AVO is not a consumer NVIDIA agent. It is a long-horizon coding-and-reasoning architecture: inspect, hypothesize, act, test, remember, revise. A supervisor watches the trajectory and redirects when the worker stalls or loops. TechCrunch called that supervisor something like a CEO over the operating agent.
Read NVIDIA’s own write-up: AVO reaches 100% on ARC-AGI-3. The honest line from that post: evaluating a model is not the same as evaluating an agent.
“The model matters, but the model is not the entire agent.”
That is NVIDIA’s close, not ours. The model is potential intelligence. The harness is what converts it into dependable work.
Hold the 30 → 100 number honestly
This is a proof point, not a magic 70-point guarantee. NVIDIA is careful about that. We should be too.
Complete agent vs raw model
ARC Prize reports ~30% for Claude Opus 5 at High reasoning. AVO + Opus 5 scored 100.00 RHAE on the 25-environment public set, 183/183 levels, in 6,624 environment actions.
A controlled ablation
NVIDIA says the comparison is not a clean “add AVO, gain 70 points.” Memory, context, observation format, and execution all differ. Public set only, not the private competition sets.
Seven days, 500+ directions
On GPU-kernel work, AVO ran seven days, explored 500+ directions, committed 40 kernel versions, and beat FlashAttention-4 by up to 10.5% in NVIDIA’s tested DGX B200 configs.
Architecture extracts usable performance
A powerful model inside a weak system produces impressive demos. A designed harness turns the same model into something that can finish long assignments.
What the harness actually is
The model is the reasoning engine. The harness is the operating system wrapped around it.
What the model sees
Not the whole universe. The right slice. Observation format, history, and what stays out of the prompt decide whether the agent can think at all.
What it can actually do
Inspect, edit, run, test, query. Without tools, a model can only talk. With tools, it can act, and be checked.
How actions get verified
AVO’s loop is inspect → plan → act → run a real test → read the result. Feedback is grounded, not hoped for.
What survives the window
Prior implementations, test results, failed approaches, useful discoveries. Resume from accumulated work instead of reconstructing the problem every turn.
Who catches the stall
A separate watcher looks at the broader trajectory. If the worker repeats itself or chases a dead end, it gets redirected.
When to stop, or escalate
Completion criteria, failure detection, and a path for another agent or a human to intervene. Long-horizon work needs an exit, not just a loop.
NOOA: the agent as ordinary software
A related NVIDIA research preview treats an agent like a Python class: methods, fields, docstrings, types. Same model. Different harness. Different results.
Typed inputs and outputs
Agentic calls have contracts, not free text. Invalid returns fail in software, not in a polite paragraph.
Pass large data by reference
The model sees a bounded preview. The full object stays live in Python. That is how they cut token waste.
Code as the action
Deterministic Python handles rules that should not depend on model judgment. The model writes the parts that need judgment.
Programmable loops
Orchestration is ordinary Python, writable by developers and, when useful, by the model itself.
Explicit persistent state
Durable typed state lives on the object, not only in chat history. Inspectable. Backup-able. Reviewable.
The agent manages its own context
Context blocks and event history are APIs the model can inspect, instead of a hidden window that silently overflows.
Efficiency, not just score
NVIDIA reports NOOA on SWE-bench Verified at 82.2% with GPT-5.5 using ~1.1M tokens per task, roughly half the tokens of comparison harnesses that landed in the high 70s. On ARC-AGI-3, NOOA is a different system than AVO: 85.1% mean RHAE with GPT-5.6-sol under $20/game. Do not mash those two scoreboards together.
Source: Six Agent Harness Capabilities.
It also supports model routing
NVIDIA’s broader stack does not assume one expensive model should do everything. NeMo Switchyard routes steps by capability, cost, latency, and escalation.
Small model, cheap step
Classification, cleanup, and mechanical follow-through do not need a frontier bill. Send them to a specialist or a local model.
Frontier only when it earns it
Strategic reasoning, stuck recoveries, and high-stakes judgment escalate. The rest of the workflow stays on the gas tank you can afford.
Validation can be code
Some gates should not be prompts. Deterministic checks decide pass/fail. The model does not get to talk its way through a failed test.
This is the AI Persona Method, in NVIDIA’s language
A useful AI Persona is not a prompt taped to a model. It is identity, knowledge, skills, tools, memory, workflows, guardrails, evaluation, and human accountability assembled into an operating system.
Brand Bible → identity
Stable judgment. The persona does not drift because the window got long. That is NVIDIA’s “explicit state,” applied to taste and voice.
Skills & SOPs → procedures
Repeatable operating procedures. The worker does not invent a new process every Tuesday.
Tools → capacity to act
The agent can do the job, not just describe it. Same lesson as AVO’s test-and-revise loop.
Memory → continuity
The business is remembered. Failed approaches stay failed. Useful discoveries survive the next session.
Supervisor + human
The agent does bounded production. A human operator reviews. The founder decides. Specialists validate the technical bits. Accountability stays human.
Gates + routing
Tests and approvals before risky actions. Cheaper models on routine steps. That is hybrid by design: humans + AI, not AI instead of the VA.
The better purchasing questions
If you are still asking “which model is smartest,” you are shopping for a demo. Ask these instead.
Does it remember the business?
And does it know which sources are authoritative? A clever answer from the wrong file is still a miss.
Can it use the required tools?
If the work lives in a CRM, a repo, or a design system, talk is not a capability.
Are outputs structurally validated?
Can it recognize failure? Or does it keep smiling while the work quietly rot?
Is there a supervisor?
Someone, agent or human, has to catch loops, dead ends, and “almost done” that is not done.
Are risky actions gated?
Can the work be audited and rolled back? If not, you do not have an employee. You have a dare.
Can cheaper models do the routine?
And does the system get better from completed work? If every step burns the frontier model, the architecture is unfinished.
This is / this is not
Boundaries keep the page honest and the hybrid story clean.
A field note on architecture
- NVIDIA research that the harness extracts usable performance
- A map from AVO / NOOA / Switchyard onto the AI Persona Method
- A buying filter for founders still shopping models
- Hybrid by design: humans + AI, silicon underneath
- A companion to our earlier harness-engineering brief
A NVIDIA product launch
- Not a new consumer NVIDIA chatbot
- Not “architecture always adds 70 points”
- Not a claim that we ran ARC-AGI-3 ourselves
- Not “AI replaces the VA”
- Not a rewrite of the fine-tuning-vs-harness page
How this sits with the older pages
Canonical “harness first” story for August 2026: this page, at /nvidia-avo-harness/.
/before-fine-tuning-harness-engineering/ stays the earlier case for evaluation before training. /ai-persona-method-update/ stays the method itself. /ai-fleet-router/ is the control plane we actually run.
If you landed on the older harness page first: since then, NVIDIA published AVO at 100% on the ARC-AGI-3 public set and put a supervisor in the loop. Start here for the current proof point.
Related reading
More from the lab if you want the surrounding architecture story.
Sources
Primary NVIDIA posts and the TechCrunch report. Dated, linked, not paraphrased into extra numbers.
Want an AI Employee, not another model tab?
The goal was never the smartest chat window. The goal is an employee that remembers the business, uses the tools, fails loud, and still leaves human judgment in charge.
