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
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 — 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.
This is a proof point, not a magic 70-point guarantee. NVIDIA is careful about that. We should be too.
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.
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.
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.
A powerful model inside a weak system produces impressive demos. A designed harness turns the same model into something that can finish long assignments.
The model is the reasoning engine. The harness is the operating system wrapped around it.
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.
Inspect, edit, run, test, query. Without tools, a model can only talk. With tools, it can act — and be checked.
AVO’s loop is inspect → plan → act → run a real test → read the result. Feedback is grounded, not hoped for.
Prior implementations, test results, failed approaches, useful discoveries. Resume from accumulated work instead of reconstructing the problem every turn.
A separate watcher looks at the broader trajectory. If the worker repeats itself or chases a dead end, it gets redirected.
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.
A related NVIDIA research preview treats an agent like a Python class — methods, fields, docstrings, types. Same model. Different harness. Different results.
Agentic calls have contracts, not free text. Invalid returns fail in software, not in a polite paragraph.
The model sees a bounded preview. The full object stays live in Python. That is how they cut token waste.
Deterministic Python handles rules that should not depend on model judgment. The model writes the parts that need judgment.
Orchestration is ordinary Python — writable by developers and, when useful, by the model itself.
Durable typed state lives on the object, not only in chat history. Inspectable. Backup-able. Reviewable.
Context blocks and event history are APIs the model can inspect — instead of a hidden window that silently overflows.
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.
NVIDIA’s broader stack does not assume one expensive model should do everything. NeMo Switchyard routes steps by capability, cost, latency, and escalation.
Classification, cleanup, and mechanical follow-through do not need a frontier bill. Send them to a specialist or a local model.
Strategic reasoning, stuck recoveries, and high-stakes judgment escalate. The rest of the workflow stays on the gas tank you can afford.
Some gates should not be prompts. Deterministic checks decide pass/fail. The model does not get to talk its way through a failed test.
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.
Stable judgment. The persona does not drift because the window got long. That is NVIDIA’s “explicit state,” applied to taste and voice.
Repeatable operating procedures. The worker does not invent a new process every Tuesday.
The agent can do the job, not just describe it. Same lesson as AVO’s test-and-revise loop.
The business is remembered. Failed approaches stay failed. Useful discoveries survive the next session.
The agent does bounded production. A human operator reviews. The founder decides. Specialists validate the technical bits. Accountability stays human.
Tests and approvals before risky actions. Cheaper models on routine steps. That is hybrid by design — humans + AI, not AI instead of the VA.
If you are still asking “which model is smartest,” you are shopping for a demo. Ask these instead.
And does it know which sources are authoritative? A clever answer from the wrong file is still a miss.
If the work lives in a CRM, a repo, or a design system, talk is not a capability.
Can it recognize failure? Or does it keep smiling while the work quietly rot?
Someone — agent or human — has to catch loops, dead ends, and “almost done” that is not done.
Can the work be audited and rolled back? If not, you do not have an employee. You have a dare.
And does the system get better from completed work? If every step burns the frontier model, the architecture is unfinished.
Boundaries keep the page honest and the hybrid story clean.
Canonical “harness first” story for August 2026: this page — /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.
More from the lab if you want the surrounding architecture story.
Primary NVIDIA posts and the TechCrunch report. Dated, linked, not paraphrased into extra numbers.
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.
Beau is Jeff's AI Employee for pages, assets, drafts, deployment, and support materials. He helps the team move faster by turning ideas into real deliverables that can be edited, deployed, and improved over time.
