Four of today’s five stories are about the software wrapped around the model rather than the model itself, which is the shift I have been arguing for since I moved this entire content operation onto a harness instead of a chat window. The fifth is a startup asking you to take its word for a benchmark it ran on itself.
Nvidia Scored 100 on ARC-AGI-3 by Changing Everything Except the Model
Nvidia’s research team reported that its Agentic Variation Operators system posted a 100.00 RHAE score across all 25 environments of the ARC-AGI-3 public set, clearing all 183 levels, and Nvidia’s own technical writeup names Claude Opus 5 as the model doing the reasoning inside it. That is not the number that traveled. What traveled was the ratio: Opus 5 manages roughly 30% on the same benchmark by itself, so the wrapper appears to have more than tripled the model, and TechCrunch framed it as the harness becoming the real hero.
The direction is right. The arithmetic is being handled carelessly, including by people who should know better.
Two things keep that 30-to-100 jump from being the clean before-and-after it reads as. The 100.00 is an RHAE score on the 25-environment public set, while the ~30% baseline comes from runs the ARC Prize foundation administers itself, so the two numbers are not the same measurement of the same thing. And Nvidia attaches an explicit warning to its efficiency comparison against VISTA, the previous best harness, where AVO finished in 6,624 environment actions against 7,542:
This should not be interpreted as a controlled ablation: the two systems differ in agent backend, observation representation, memory, context management, and other implementation details.
That is Nvidia telling you, in its own post, that cross-system comparisons here are directional rather than decisive. Nobody quoting the tripling put that sentence next to it. The semi-private and private held-out sets remain untouched, and those are the ones that historically separate a real result from a well-tuned one.
Verdict: matters, and the weaker headline is the stronger story. Bury the benchmark and look at the kernel work instead, where the same system explored more than 500 optimization directions over seven days, committed 40 kernel versions, and beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on DGX B200 hardware. That is a real task with a real baseline that Nvidia has every incentive to measure honestly, because it ships the hardware. A supervisor process that notices an agent has stalled and redirects it is the single highest-leverage thing most people are missing, and it is not a model upgrade. It is about forty lines of orchestration.
AWS Added a Domain Allowlist to Agent Web Search, and No, It Did Not Just Go GA
Half the coverage I read this weekend said AWS pushed Web Search on Bedrock AgentCore to general availability on August 21. It did not. AWS shipped GA on June 16, in US East (N. Virginia) only. What actually landed last week was better than a GA flag anyway.
The August 19 update gives agents per-call domain filtering, so a single request can restrict itself to trusted sources or block unwanted ones without an admin touching the configuration, plus published-date filtering with inclusive from and to bounds. Administrators get gateway-level allowlists holding up to 100 domains. It also expanded past Virginia into Europe (Ireland) and Asia Pacific (Tokyo), which is the part that decides whether a European client can use it at all.
Note what week this is. Anthropic shipped the same primitive days earlier, and I wrote up its allowed_domains and blocked_domains controls in Thursday’s roundup as the sleeper of that release. Two of the three biggest agent platforms independently concluded within days of each other that the fetch tool needs a fence enforced in the API rather than a polite request in a prompt. When competitors converge that fast, they are responding to the same support tickets.
Verdict: matters. Every research agent I have built fails the same way, which is that it finds a confident-sounding aggregator, treats it as a source, and launders a wrong number into a draft. I hit exactly that failure while researching this post, where three separate sites reported an August GA date that never happened. Date filtering is the underrated half: most of my bad retrievals are not untrustworthy domains, they are trustworthy domains serving a 2024 page as though it were this week’s.
A2A Moved Into the Agentic AI Foundation, So MCP and A2A Now Share a Landlord
The Agent2Agent protocol joined the Agentic AI Foundation as a hosted project, and the foundation’s announcement puts more than 150 organizations behind A2A, with AWS, Cisco, Google, Microsoft, Salesforce, SAP and ServiceNow as founding members. AAIF now hosts the layers as a set: AGENTS.md for instructions, goose for the runtime, MCP for agent-to-tool, agentgateway for traffic control, and A2A for agent-to-agent.
Verdict: matters, quietly, and mostly as risk reduction. The thing that has made me slow to build on any agent protocol is not technical, it is governance. A protocol controlled by one vendor is a protocol that can be repriced or reoriented the moment it stops serving that vendor’s roadmap. MCP and A2A sitting under the same neutral foundation does not make either better engineering, but it changes the shape of the bet. I would still not build a business-critical path on A2A this quarter, because a hosted project is a governance change and not a maturity signal. I would stop treating it as a Google protocol, which is what I was doing.
Inherent Says a 27B Model Beat Opus 4.8 and GPT-5.5. Inherent Also Ran the Test.
Inherent, founded by DeepMind alumni, says its agent Faraday outperformed frontier systems at independently reproducing the findings of published scientific papers. TechCrunch’s writeup puts Faraday on Qwen 3.6 at 27 billion parameters, measured against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5.
Here is everything the announcement does not contain: who administered the evaluation, the size or composition of the test set, the accuracy metric, and any external party willing to attach its name to the result. The comparison appears to have been run by the company that benefits from it. To Inherent’s credit, co-founder Edward Hughes says the methodology matters more than the result, which is the most defensible sentence in the whole story and also an admission that the number is not the point.
Verdict: marketing, in the specific sense that it is a real research direction wearing a benchmark claim it has not earned. I flagged the same pattern with vendor agent-security statistics a week ago, and the tell is identical: a striking ratio, a self-run evaluation, and no test set you can inspect. Small models beating large ones on narrow tasks is genuinely happening and I would love this to hold up. Reproducing a published paper is also a task where the answer is in the training data, which is the hardest possible case to prove you did it honestly. Wait for someone else to run it.
The Assistants API Dies Tuesday
Short one, because I covered the substance on Thursday and the only new fact is the clock. OpenAI’s deprecation schedule puts the Assistants API sunset at August 26, which is 48 hours out. Calls to the assistants, threads and runs endpoints stop working, and OpenAI is not migrating your threads for you.
Verdict: breaks your stack, and it is the last morning that sentence is useful to you. If you have a live Assistants integration and have not cut over to Responses and Conversations, you are no longer planning a migration, you are planning an incident. The realistic move now is not a clean port. It is to find every call site today, stand up the smallest possible shim, and buy yourself the week you should have taken in July. The full breakdown from Thursday has the endpoint mapping.
What I Am Watching
The harness story and the domain-allowlist story are the same story told twice, and it is the one worth internalizing: the differentiated thing you own is increasingly the scaffolding, not the weights you rent. On the marketing side, watch whether anyone independently reproduces Inherent’s replication claim, because a self-run benchmark that nobody repeats is a press release with a chart. And if the Assistants API takes your product down on Tuesday, the postmortem does not say OpenAI broke it. It says you had twelve months.