Hey there 👋 I kept waiting this week for the one story. Some mornings it's a new model that resets the benchmarks. Some mornings it's a nine-figure round that tells you where the money thinks all of this is going. I opened tab after tab looking for that single headline and never found it. What turned up instead was three different labs shipping three different pieces of the same machine. NVIDIA put out a way to write an agent as plain code. Google flipped the switch on the enterprise plumbing that gives an agent an identity and an audit trail. OpenAI shipped the operating layer that keeps a production agent inside its lane. None of them is the story on its own. Together they are: the stack you need to run an agent inside a real company went from a pile of homegrown parts to something you can mostly buy. So today is a build-it issue, not a watch-it one. Below is what landed, and which layer each piece fills. Ron
The Big ThingNVIDIA made an AI agent a single Python class.NVIDIA Labs open-sourced NOOA, short for NVIDIA Object-Oriented Agents, and the pitch fits in one sentence: your whole agent is a single Python class. Not a prompt file sitting next to a tool schema sitting next to a callback module sitting next to a workflow graph. One class, read top to bottom like any other object. The mapping is the clever part. The class methods are the actions the model is allowed to take. The fields are the agent's state. The docstrings are the prompt. The type annotations are contracts the runtime enforces, so a function that promises an integer has to hand one back. Leave a method's body as just ... and it gets filled in at runtime by an LLM loop. Write the body normally and it stays plain, deterministic Python. One file ends up holding both the reasoning and the hard-coded logic, which is a very different thing from stitching four systems together and praying they agree. The models are pluggable through LiteLLM, so a hosted API, a local Ollama model, or a self-hosted vLLM endpoint all drop in behind the same code. It's Apache 2.0, it installs with pip install nooa, and NVIDIA is pointing at benchmark numbers to back the design: 82.2% on SWE-bench Verified, 86.8% on CyberGym L1, and 85.1% mean on ARC-AGI-3. Grade those the way you grade any launch-day scorecard. They're NVIDIA's own, run on NVIDIA's harness, and nobody outside has reproduced them. Trust the mechanism first and the percentages later. Ship it? Prototype yes, production no, and the version number tells you why. PyPI marks the current release, v0.0.8 from July 30, as alpha. NVIDIA itself calls the project a research preview, and it only runs on Python 3.12 and 3.13. That is not what you put in front of paying customers this quarter. It is a great use of one Friday afternoon: clone it, port a single small agent over, and see whether the class reads cleaner than the graph it replaces. If the object-oriented model clicks for your team, it deletes a whole category of glue code you're maintaining by hand right now. NVIDIA NOOA · MarkTechPost
Tour de Headlines🔑 Google's enterprise agent control plane went generally available. Google Cloud moved the operational half of its Gemini Enterprise Agent Platform to GA in one go: Agent Runtime, Memory Bank, Gateway, Registry, Evaluation, Observability, and the one worth your attention, Agent Identity. It's a native IAM type built on open standards that gives an agent least-privilege permissions, binds its access to the runtime so a stolen token is useless by itself, writes a non-repudiable log of every action the agent takes, and retires credentials automatically so you stop leaving dormant keys around. If you've been faking agent identity with a shared service account, this is the version with a real name attached. The customers Google names, AT&T, Best Buy, Commerzbank, Liberty Global, and WellSky, are its own list and mix live users with evaluators, so confirm which specific services are GA in your region before you build against them. Google Cloud ⚙️ OpenAI Presence is the boring operating layer, and that's the point. Presence bundles what it takes to hold a production agent to its job: written policies and standard procedures, permissions and guardrails, a fixed set of approved actions, simulations to test weird inputs before launch, and a Codex-powered loop that proposes improvements a human signs off on before they ship. Every deployment starts from one business task and the agent gets only the access that task needs. The proof point is OpenAI's own phone support, which now runs on Presence and closes 75% of inbound calls without a human, with the Codex loop trimming handoffs by 15 points over ten days. It arrived in late July, so it's the layer that showed up just ahead of this week's build and identity pieces. Those numbers are OpenAI's own desk grading itself, and there's no public pricing yet. OpenAI 🤖 Customer ops is where agents are getting hired for real. Berlin's telli raised a $15M seed led by redalpine, with Cherry Ventures and Y Combinator in, to run customer-facing operations end to end: voice calls, WhatsApp, SMS, live chat, and email through a single system. At the center is an AI coworker named Charlie that helps ops teams build and tune the agents, wire in internal systems, and read back what's going wrong in conversations. Clients including Sky, Viessmann, and Vaillant already push millions of customer interactions through it. This is a funding round, not a benchmark, so weigh it as a read on demand rather than a product you can score. The read is loud: the layer sitting on top of all this plumbing, agents doing real customer work, has paying enterprise buyers today. TechFundingNews
Sponsor Read the human signal in every conversation. You instrument your agents. RapportScore does it for the humans, reading the communication signals in live sales and support calls so your team can see what landed and adjust before the next one. Built by the crew behind this newsletter. See your team's score → |
Tool of the Day🖥️ Gemini CLIWhat it is for: putting a real agent on a real task from your terminal this afternoon, for free. If NOOA is the framework you're intrigued by and can't ship yet, Gemini CLI is the one you can use before lunch. It's Google's open-source terminal agent: point it at a repo and it reads the code, calls local and remote tools, and speaks MCP, so it plugs into the same servers your other agents already use. It runs from the command line, it's free to start, and it turns "I should try an agent on this" into a five-minute experiment instead of a standing project. This isn't fresh news, it's the easy option that's been sitting there the whole time everyone benchmarks the frontier. Capability tracks whichever Gemini model tier you point it at, so a heavier task wants a heavier model behind it. Get Gemini CLI on GitHub
Worth a Click- AMD bought its way into Anthropic's compute. AMD said it will invest up to $5 billion in Anthropic, and Anthropic will deploy up to 2 gigawatts of AMD's MI450 Instinct chips plus its Helios rack systems, with the first gigawatt landing in the first half of 2027. It was announced at AMD's Advancing AI event in late July. The money is a committed future equity stake, not cash already wired, and both the dollars and the gigawatts carry an "up to." Read it as Nvidia finally getting a serious second supplier for the compute your agents run on. CNBC
- Microsoft picked Go for its agent framework. Per The New Stack, Microsoft's Agent Framework added Go support, putting Microsoft alongside Google in betting on Go for agent development while OpenAI and Anthropic stay Python and TypeScript first. It's one outlet's framing and a small thing today, but the language a big vendor standardizes on is a decent leading indicator of where production agent code heads next. The New Stack
Delight👏 The robot trained for millions of hours to carry one mini-fridge. Boston Dynamics had its Atlas robot walk over to a fully loaded mini-fridge, pick it up, and carry it across the lab with nobody at the controls. The good part isn't the lift. It's that Atlas had never seen that fridge in training. It practiced on lighter objects for millions of parallel hours in simulation, with the engineers randomizing floor friction, motor strength, and how heavy things were, until hoisting a strange heavy object it had never met worked on the first try. Sit with that for a second, because it's the whole issue in one image. The clip that goes viral is three seconds of a robot bear-hugging a fridge. The work that made it reliable was millions of boring hours nobody will ever watch. Every software agent that survives contact with real users got there the same unglamorous way.
The agent production stack filled in this week.
Put the pieces next to each other and they line up into one machine. NVIDIA gave you a way to write the agent as plain code. Google gave the agent an identity, a permission boundary, and an audit log that holds up in a review. OpenAI gave you the operating layer that keeps a live agent on its policies and improves it under human sign-off. telli is the proof that companies will pay for the finished result. A year ago you built every one of those yourself, badly, and taped them together with a scheduler and hope.
So the move for anyone running agents past the demo is to stop treating all of it as one giant homegrown project. It's three layers now, build, govern, operate, and each one has real options you can adopt instead of maintain. Find the layer you're weakest at. If your agents have no identity and no audit trail, that's your Monday. If they run but you can't test or constrain them, start at the operating layer. The teams that win the next year are the ones who stop hand-rolling the parts that just became infrastructure, and spend the time they save on the part that's still theirs to get right: the job the agent is there to do. Ron |
