Daily Briefing
28 August 2026 · 5 stories
🔥 Top story
1AccuKnox launches AgentZ: the first platform that makes AI agents truly enterprise-ready
Today, most AI agents are impressive in demos but stuck in production: security teams refuse to let them touch customer data. AccuKnox, based in Menlo Park, launches AgentZ on August 27: a platform that bundles into one product the agent itself, its isolated sandbox, its tools, its permissions, and its governance. CEO Nat Natraj puts it bluntly: "Organisations have moved past the demo stage. They need structure, not another framework." Concretely, a company can now deploy the same agent on its cloud, on its own premises, or completely disconnected from the internet — depending on what the regulator demands. This is the eighth major player in eight days to adopt the same production-ready agent architecture — as if the whole industry was aligning its electrical plugs on a single standard. Source: https://www.accuknox.com/agentz
2Sequel turns its webinars into an AI agent that knows every attendee by name
If you've ever attended a corporate webinar, you know the video lands in a drawer once it's over. Sequel, which runs webinars for businesses, flips the script on August 26: its platform becomes an "Agentic Engagement Cloud" where three ready-made agents (Build, Scale, Convert) analyse who watched, who dropped off, who asked questions — and write the personalised follow-up that the sales team should have sent. An MCP server exposes that same data to Claude, ChatGPT, or any external agent, like a universal USB plug for customer engagement. Sequel's strategic message: "the advantage isn't a better agent, it's what the agent knows." For an SME that runs webinars, one thing changes concretely: the post-event recap writes itself and can tell the curious attendee from the hot prospect. Source: https://www.sequel.io/news/introducing-agentic-engagement-cloud
3GLM-5.3, the Chinese giant with 743 billion parameters, opens its weights today on Hugging Face
The countdown is on: on the Hugging Face page of lab Z.ai, you can read "Expected release August 28, 2026" and 83 researchers are following the release live. GLM-5.3 is a monster of 743 billion parameters — imagine 743 trillion internal settings that the AI adjusts to understand language — with 40 billion active at each pass. It can read one million tokens at once (the equivalent of a 700-page novel) and remains the best open-source model on Terminal-Bench, the test that measures a model's ability to run command-line tasks like a developer would. If Z.ai publishes the weights under an MIT licence — truly free and reusable — then any company can install this model on its own servers. If the licence is more restrictive, it confirms a trend: China keeps its best models under control for a few weeks after release. Source: https://huggingface.co/zai-org/GLM-5.3
4Runable raises $21 million to sell a general-purpose AI agent to small businesses
While giants fight over trillion-parameter agents, an Indian startup in Bengaluru proves there's a huge market at the bottom of the ladder. Runable, founded in 2026, announces on August 26 a $21 million Series A and already claims $2 million in recurring revenue in three weeks, with one and a half million users. Its target: very small businesses (agencies, consultancies, cleaning companies) that don't have a technical team but need an assistant to handle schedules, quotes, and follow-ups. The CEO says it simply: "Nobody starts a business to have a nice landing page. You start it to get customers and revenue." The agent installs without a developer, speaks English or Japanese, and starts working from the first login. For an SMB juggling three tools and a spreadsheet, it's the promise of replacing the entire software stack with a single assistant. Source: https://yourstory.com/2026/08/runable-raises-21m-series-a-ai-agent-startups-india
5Andreessen Horowitz verdict: an AI agent consumes five times more tokens than a human
Andreessen Horowitz, one of the most influential venture-capital firms in Silicon Valley, publishes on August 26 a data-driven study on an unexpected angle: the real cost of an AI agent compared with a human using ChatGPT. The result surprises: an autonomous agent burns about five times more tokens — the text units that AI providers bill for — than a human, of which 85% in hidden calls the user never sees. The cause is mechanical: where a human sends a question then waits for the answer, the agent loops endlessly to plan, use tools, verify, continue. For an SMB that charges its clients a flat fee, this means an "agent" account costs five times more than a "human" account on the same platform. It's precisely this economic calculation that drove Stripe to put $7.5 billion on the table to acquire OpenRouter, the router that distributes those requests across models. Source: https://www.eweek.com/news/a16z-ai-agents-consume-5x-more-tokens-than-humans/
📡 To watch
Will GLM-5.3 ship under an open licence or under control?
Everything hinges today: if Z.ai publishes the model weights under MIT (the most permissive open-source licence), it confirms the Chinese canon stays open and reusable by any Western company. If the licence is more restrictive, it's a signal that Beijing is hardening its doctrine and keeping its best models under control for a few weeks after release. The expected answer on the Hugging Face page by end of day will set the rhythm of the open-weight AI market for the next six months.
Will Stripe really close its $7.5 billion acquisition of OpenRouter?
According to the New York Times and CNBC, Stripe offered $7.5 billion to absorb OpenRouter, the service that routes AI requests to the best available model. The investor letter from August 19 mentions a closing "in the coming weeks". Any change in price, terms, or timeline will be read by the market as a signal of the real value of agentic infrastructure — the layer that doesn't build the model, but distributes its answers.
Will the major robotics labs adopt the Chinese benchmark RoboColiseum?
Officially launched on August 24 in Shanghai, the RoboColiseum platform offers four evaluation dimensions across 78 high-fidelity robotics tasks, with a sim-to-real gap below 10%. Hundreds of teams are already in closed beta. If American players like Figure, Tesla Optimus, or Unitree join the platform, it creates the first global evaluation standard for embodied AI — the equivalent, for robots, of what the HELM or MMLU tests did for text models in 2020-2021.
Multilogin releases an open-source agent that pilots real Android phones in the cloud
The Dubai-based startup published on August 26 an open-source agent that connects AIs to real Android phones hosted in the cloud. Concretely, the agent sees the screen, types text, opens native apps — what browser-based bots can no longer do, since platforms like TikTok, Instagram, or X detect and block them. For an SMB managing twenty social accounts, this opens the door to automation that doesn't get kicked out every two hours.
📊 Trend
In eight days, eight major players — Snowflake, AWS, Google, NVIDIA, Meta, Okta, Salesforce, AccuKnox — agreed, without consulting each other, on the same enterprise AI agent architecture: an isolated sandbox, observable tools, clear governance, and now the ability to deploy the same agent in the cloud, on premises, or fully offline. Meanwhile, the Chinese market proves it can deliver a 743-billion-parameter model in a matter of weeks, and an Indian SMB signs $2 million in revenue in three weeks with a general-purpose agent that talks to two-person companies. The AI agent is no longer a lab experiment: it's now a commercial product that you buy, deploy, and bill by usage — with its own economics, where each agent counts as five human users on the invoice.