Gemini Agents to Do Office Jobs and Amazon Invests More in AI Than Entire Countries
- • Google presents the Gemini agent as an innovative colleague
- • Amazon surpasses the German state with record investments in AI
- • OpenAI corrects revenue forecasts, figures far below expectations
Google Turns Gemini Agents into Colleagues
At its Gemini at Work 2026 event, Google Cloud introduced the “Gemini agent,” a single universal agent for enterprise work. It was presented by Google Cloud CEO Thomas Kurian. The system is designed to handle research tasks, document creation, code, and the coordination of other agents, with tasks potentially running for hours or days. Employees are meant to delegate goals instead of giving individual instructions; the agent plans the steps itself and returns a finished result. The product is currently in a private preview for enterprise customers, with broad availability planned for selected Workspace Business and Enterprise plans.
The agent runs in the cloud and, according to the provider, maintains a single memory and context state across all devices and channels. It is accessible via web, iOS, Android, Windows, Mac, command line, Google Workspace, Microsoft 365, and Slack, and can also run in third-party applications without its own interface. Within Workspace, it works directly in Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar. According to the company, it organizes its knowledge into four types of memory: session memory for ongoing tasks, semantic memory from documents and conversations, procedural memory for workflows, and episodic memory for completed jobs. For complex tasks, it can form temporary groups of specialized sub-agents.
The announcement gets particularly specific when it comes to identity. A manager can, for example, create an “Event Planner Agent” that gets its own Workspace account, its own email address at @agents.company.com, a calendar, Drive storage, and an entry in the company directory. Employees are supposed to assign work to it like a colleague, for example, by mentioning it in Google Chat. According to Google, such an agent only sees the information the team shares with it and does not automatically inherit broader organizational permissions. For security, the company mentions identity and policy management, entitlement controls, sandboxed execution environments, and network gateways.
The model architecture is open: In addition to Google’s own Gemini models, the agent already uses Claude models from Anthropic, with more proprietary and open models to follow. A feature called Smart Routing evaluates each task and assigns it to the model with the best quality-to-cost ratio; administrators can set spending limits at the project level. For context, Kurian mentioned that last year, nearly 500 Google Cloud customers each processed more than a trillion tokens, and about 90 percent of the Fortune 100 use Gemini Enterprise.
The announcement is one in a series of similar moves: Microsoft introduced a revised Copilot on September 25, OpenAI its continuously running 'dots' agents four days later, and Anthropic a native Claude integration with Google Docs, Sheets, and Slides on October 6. xAI had launched its persistent Grok-Bot agents in August and expanded them with shared Team Bots in September. → Google Cloud Blog, VentureBeat, The Verge, Quartz
Synthszr Take: An agent with its own address at @agents.company.com sends emails that must be legally attributed to someone, and that someone is not the agent. Google is building the identity layer in a technically clean way: its own account, its own permissions, only the data the team shares. This doesn’t answer the liability question, because the person who created the Event Planner Agent has to take responsibility, and after three days of it running across four types of memory, they probably no longer know on what basis a decision was made. Therefore, every agent account needs a named owner in the company directory and a log that makes every action traceable before the first agent goes live.
In 2026, Amazon Is Investing More in Data Centers Than the Entire German State
This year, Amazon will invest around $222 billion in data centers for artificial intelligence, according to a projection by Bloomberg. For comparison, the Handelsblatt AI-Briefing cites Germany’s gross fixed capital formation, which amounted to the equivalent of $170.6 billion in 2025 and covers roads, schools, and hospitals. The orders of magnitude are also converging on the revenue side: Amazon’s revenue in 2025 was about $718 billion, and for this year, analysts surveyed by Bloomberg expect $829 billion, while Japan recorded tax revenues of $809 billion in 2024. In terms of creditworthiness, Microsoft, with a Aaa rating, is one step above the USA, which is at Aa1; Meta ranks on par with France and the United Kingdom. → Handelsblatt KI-Briefing
Synthszr Take: $222 billion in data centers is infrastructure policy without a parliament. When a single corporation builds more in twelve months than a G7 state does in roads, schools, and clinics, that corporation decides where electricity, grid connections, and construction sites go, long before a budget committee even hears about it. The credit ratings make this shift even more tangible: Microsoft gets capital on better terms than the United States, and Amazon can sustain investment cycles where a finance minister would have had to pull the plug long ago. Antitrust law is the right framework, but a framework is only effective if Brussels also enforces it against Washington, and the signals of recent months suggest capitulation rather than enforcement.
OpenAI’s Revenue Significantly Lower Than Previously Reported
OpenAI has reported an annualized revenue of nearly $50 billion to its investors, reports The Information. This figure is below the numbers that were previously circulating in reports about the company’s revenue pace. The metric is an Annualized Revenue Run Rate: The most recent monthly revenue is extrapolated over twelve months and does not correspond to the actual booked annual revenue. Such figures are provided to existing and potential investors and are not publicly audited. → The Information
Synthszr Take: Nearly $50 billion annualized is a huge number, and yet it’s still below what was reported for months. This gap arises from the chain of circulated target figures, passed-on forecasts, and the number that ultimately appears in the investor deck. For a company that backs its funding rounds and data center commitments with expected growth, the reported number is itself a tool: it finances the infrastructure before the revenue can support it.
Atlassian Introduces New AI Features for Jira & Co.
At its Team 26 Europe event, Atlassian introduced new AI features for Jira, Confluence, Loom, DX, and Jira Service Management that link agent work to its in-house Teamwork Graph. This is based on a previously published playbook for an AI-native SDLC, which redefines planning, design, development, review, and maintenance instead of just layering AI onto the old process. The impetus comes from Atlassian’s own 2026 study: 94 percent of engineering leaders surveyed say their teams are already using AI, but only 6 percent feel they are equipped with systems to scale and manage it. The Teamwork Graph connects code, documents, decisions, and over 80 third-party sources in a layer that understands permissions. According to the provider, internal tests with this context layer resulted in up to 44 percent better response quality with up to 48 percent fewer tokens. → AI Secret
Synthszr Take: 94 percent use, 6 percent manage: this 88-point gap is the most interesting finding of the entire release. Agents that review Pull Requests and assist with incidents are making decisions that a named individual was previously responsible for, and the speed of these decisions is no longer tied to any approval process. Atlassian’s answer is a context layer that understands permissions, which is the sensible lever to pull: an agent can’t argue away an access rule, but it can argue away a guideline in Confluence. The 44 percent better response quality comes from the provider’s internal tests and says nothing about which decisions an agent is allowed to make alone and where a human must sign off.
OpenBot Lets Codex, Claude, and Gemini Work as an Agent Team on Your Own Computer
OpenBot is a free desktop application that allows multiple AI agents to run in parallel on your own computer and hand off tasks to each other. Each agent gets its own name, its own instructions, and its own workspace; messages, files, and handoffs are exchanged between them in shared channels. It uses existing subscriptions and tools as its foundation: Codex via the ChatGPT plan, Claude Code via the Claude plan, Gemini via Google AI Pro or Ultra, plus Grok, OpenCode, Cursor, and Cline. Alternatively, any OpenAI-compatible endpoint or a local model via Ollama and LM Studio can be connected. According to the provider, workspaces, conversations, files, and browser data remain on the computer running OpenBot. → Techpresso
Synthszr Take: Zero dollars for the layer that coordinates agents and hands off tasks between them. A year ago, this was the exact product promise for which startups raised funding rounds; today, it’s available as a download for macOS, Windows, and Linux and earns nothing itself. When seven providers from Codex to Cline are interchangeable behind the same agent window, the scarcity shifts to the formulation of the task: What should the agent do, and how does someone know it has done it wrong?
Nearly One in Two US Employees Exaggerates Their AI Usage
The workforce analytics company Visier found in a survey of US employees that nearly half admit to having exaggerated their own AI usage or competence to colleagues or management. Visier calls this pattern “performative AI.” 45 percent of respondents said they felt pressure to use artificial intelligence, even if they are unsure how to do so effectively. Andrea Derler, who is responsible for research and value analysis at Visier, attributes this behavior to this pressure of expectations in daily work. → Business Insider
Synthszr Take: As soon as AI usage appears on a dashboard, it becomes a number that employees serve, and the 45 percent feeling of pressure is the record of that. An organization gets what it measures: a few prompts before the status meeting, a screenshot in the team channel, and the adoption rate looks wonderful. This ruins your own data foundation, because if nearly half the workforce is talking up their own competence, none of these numbers are suitable for an investment decision on licenses, training, or personnel planning.
OpenAI Reports Teenagers Use ChatGPT for Under 15 Minutes a Day
OpenAI has stated that, according to its own data, teenagers use ChatGPT for less than 15 minutes a day on average. Reuters reports on the announcement amidst growing concern about the risks the chatbot poses to minors. The company is under pressure regarding its handling of young users, including through legal disputes and inquiries from regulatory authorities. The usage duration is self-reported by the provider, and there is no independent verification of the data collection method. → Reuters
Synthszr Take: Fifteen minutes a day is a number that no one outside of OpenAI can verify. It comes from their own logs, with their own definition of who counts as a teenager, their own way of counting a session, and without a third party to review the methodology. A company that until now could only estimate its users' ages suddenly provides a precise average for this exact age group: the contradiction is right there in the first sentence of the announcement. Averages are the kindest tool in statistics because they make the small group invisible—the very group that the lawsuits and hearings are about, namely the teenagers having hours-long conversations at night.
Mathocalypse: Mathematicians' Association Calls for Break with OpenAI After 722 AI Proofs Are Released at Once
On October 6, OpenAI posted 722 papers with machine-generated mathematical results to a public GitHub repository, grouped into 372 results. They were generated by an internal model that is not available to outsiders; according to the company, it processed around 4,000 problems, averaging the equivalent of about three hours of ChatGPT Pro thinking time per result. The collection ranges from number theory to algebraic geometry, and according to reports, individual results concern three of the still-unsolved Millennium Problems, including the Riemann Hypothesis. Many proofs are provided with Lean formalizations, along with ten summaries of the chain of thought and data on computational effort. OpenAI acknowledges that non-formally verified proofs may contain errors and commits to fixing reported issues.
The release was made in consultation with the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, whose members include Timothy Gowers and Edward Witten and whose guidelines are based on feedback from over 600 mathematicians. OpenAI stated that it follows these recommendations with one exception: the group had requested that top-tier mathematical problems no longer be tested on internal models, and this is precisely what the company intends to continue doing.
In a statement, the Association for Human Mathematics calls on mathematicians to cease collaboration with OpenAI. It points to ongoing legal proceedings against the company for plagiarism, copyright infringement, and trademark dilution, and writes that no one in the field asked for this work. The simultaneous release of over 700 files is not a scientific achievement but a demonstration of power.
Complexity theorist Scott Aaronson reports that among the results is a proof of Subhash Khot’s Unique Games Conjecture, on which his wife, Dana Moshkovitz, worked for years. Other results concern L=BPL as well as Fourier transformation and multiplication below the O(n log n) barrier that has been in place since the 1960s. A Lean certificate exists for some of the proofs, but not for others, and according to Aaronson, no human has yet truly understood the arguments. Moshkovitz describes the texts as barely readable, with irrelevant citations and a novel, recursive code construction. A month earlier, OpenAI had processed the Navier-Stokes equation with 10,000 agents working in parallel for 88 hours.
In parallel, the release triggered a security debate in the crypto industry. On October 7, Ethereum researcher Justin Drake called for preparing a 'bunker mode' and controllably moving funds to fresh addresses whose public keys are still behind a hash. In a worst-case scenario, ECDSA could fall 'in months, not years,' by which he means the reconstruction of a private key within about a week on a large GPU cluster. Vitalik Buterin supported the concern but advised against hasty moves, explaining that botched migrations had cost him more money than all hacks combined. He extended the warning to lattice-based methods like ML-DSA and homomorphic encryption, citing this as a reason for Ethereum’s move toward hash-based signatures. Yehuda Lindell, head of cryptography at Coinbase, disagreed, saying there is no evidence that the decades-old assumptions behind elliptic curves are about to fall. → The Block, AHM, Cointelegraph, The Crypto Times, CryptoPotato, CryptoSlate, Decrypt, Bitcoin Insider, Shtetl-Optimized, Quartz
Synthszr Take: A proof used to be a proof when enough experts had followed it and found it to be correct. OpenAI replaces this with a Lean certificate and 372 result groups that no one has read yet: the machine checks the machine, and the field gets to read it afterward. The advisory group around Gowers and Witten drew exactly one red line—testing on internal models—and OpenAI has announced it will continue to cross it. That is the real statement of power, not the number of papers. The boycott call from the Association for Human Mathematics is the only leverage a community without its own data center still has, and it will fizzle out as long as the results hold. If Moshkovitz’s verdict on readability stands, a field is emerging in which correctness can be mechanically proven, and understanding becomes optional. The rules for this are being written by a lab with a model that no one from the outside is allowed to touch.
