$ today_in_10_seconds
NVIDIA published two blog posts an hour apart today. One says the harness turns a 30% model into a 100% agent. The other says the harness is the worst possible place to put your security controls. Both are correct, and together they're the clearest statement yet of where agent engineering actually stands.
$ lead_story
NVIDIA's AVO clears the entire ARC-AGI-3 public set
NVIDIA took AVO — the agent system it built to autonomously optimise CUDA kernels — pointed it at ARC-AGI-3 without changing the agent, and completed all 183 levels across all 25 public environments at a 100.00 RHAE score. The model underneath is Claude Opus 5, which ARC Prize separately verifies at roughly 30% on the same public set. Read the caveats, because NVIDIA prints them: this is the public set, not the semi-private or private competition sets, and the team explicitly says the 30%-to-100% gap should not be interpreted as a direct measurement of AVO's contribution, since the setups differ in backend, observation format, memory and context management. The genuinely interesting claim is the narrow one — the same loop that ran seven days on attention kernels (500+ optimisation directions, 40 committed versions, beating FlashAttention-4 by up to 10.5% on DGX B200) transferred untouched to a game environment with no rules, no goals and no instructions. What transferred wasn't knowledge. It was persistent memory, a supervisor that intervenes when the search stalls, and the ability to recover.
// why_it_matters: If your agent's performance is a property of your harness rather than your model, so is your agent's failure mode.
$ also_worth_attention
NVIDIA, same day: don't put your security in the harness
Published at the same hour as the AVO result, NVIDIA's AI security team argues that the harness — the exact layer that just produced the 100% — is structurally unfit to enforce security, because a layer designed to be modified cannot reliably enforce controls against its own modification. Their framing is the sharpest one-liner in agent security this year: the harness guides what an agent tries, the infrastructure determines what it can do, and only one of those is authoritative. The post maps a five-layer stack (distribution, meta-harness, harness, secure runtime, inference plane), names Claude Code, Codex and DeepSeek Harness as harness-layer examples, and lists four workload profiles from isolated to adversarial. The design rule to steal: above proposes, below decides — no model, agent, harness, tool or memory system grants itself authority.
link →The ARC-AGI-3 public set was already saturated — in July
NVIDIA isn't first to 100. The Tycho project's published scorecards show official competition-mode runs hitting 100.00 RHAE on all 25 public games with both Claude Opus 5 and GPT-5.6 Sol, under an actor-controlled world-model policy, back on 30 July. Tycho's own paper says the quiet part: with only 25 games and many of them saturating, aggregate means are not sufficient to compare systems. ARC Prize saw this coming and created a separate community leaderboard for harness results — public, self-reported, and explicitly unverified, with the foundation cautioning against reading anything into the scores. Treat every harness number you see this quarter as a self-reported one until a private-set run says otherwise.
link →Debian is voting on whether AI can touch its source at all
Debian's General Resolution on LLM-assisted contributions is live, with eight ballot options ranging from an outright ban on AI-generated and AI-assisted work to a disclosure-plus-accountability framework requiring contributors to verify licensing, take responsibility for correctness, and flag bulk AI changes in advance. GCC's steering committee already banned LLM-derived contributions on 29 July, citing the January 2025 US Copyright Office position that AI output has no copyright owner. The agentic angle is provenance: as coding agents move from suggesting lines to committing branches, who wrote this stops being a courtesy field and becomes a licensing precondition. If you ship into upstream open source, your agent's commit trail is about to be somebody's compliance artefact.
link →
$ bigger_picture
The trend behind the headlines
The two NVIDIA posts are the same finding read from opposite ends. The harness is where capability now comes from — that's the AVO result — which is precisely why it's the wrong place to put a guarantee. A layer you tune weekly for performance is a layer whose security properties you're also changing weekly, usually without noticing. That's the actual lesson of a 30-to-100 jump: the thing that produced it is programmable, and programmable means untrustworthy in the specific sense that matters to a control plane. Meanwhile Debian is arguing about the same boundary one level up — not what an agent can do, but whose name goes on the output and who carries the liability. Capability questions get answered in a quarter. Authority and provenance questions are the ones that stay open.
NVIDIA OpenShell
The secure runtime underneath today's security post, open source on GitHub. It creates the boundary before the agent launches: the harness starts inside the runtime, its plugins and MCP processes inherit the same containment, and subagents get delegated child runtimes with ceilings they can't exceed. That last detail is the one most homegrown sandboxes miss. The design principle is worth the read even if you never run it: a control the agent can decline to invoke is not a control.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime