The best AI agent is not the most autonomous one. It is the one given the right autonomy—and the right authority—for the job.
Today, we call everything an “agent,” from systems that need approval for every action to systems that plan, adapt, and execute for hours. That makes 3 basic questions hard to answer:
1. How much human supervision does this agent need? 2. What is it authorized to do? 3. Who is accountable when it acts?
The autonomous-vehicle industry faced a similar challenge. Its L1–L5 framework defines levels of automation and how responsibility shifts between driver and system. AI agents need an equivalent framework: A1–A5. Here is a proposed framework, shaped by perspectives across research, investing, and company building: @Ramesh Raskar at @Massachusetts Institute of Technology, @Vijay Reddy at @Mayfield, and @Rahul Todkar at @Nasiko:
– A1 (Directed): Performs individual actions with human approval. It may prepare a payment, but a person must authorize it.
– A2 (Delegated): Completes a bounded, multistep assignment. Tools, permissions, budget, and duration are set in advance.
– A3 (Adaptive): Revises its plan and handles routine exceptions within established limits. When it reaches the edge of its competence or authority, it stops and escalates.
– A4 (Self-Governing): Runs without routine supervision inside a certified domain. Humans set policy, define boundaries, audit outcomes, and investigate incidents.
– A5 (Independent): Operates across organizations and changing environments. Governance requires shared identity, delegation, audit, and dispute-resolution systems.
These levels are risk settings, not a maturity ladder. Higher does not mean better, and A5 should not be the goal for every agent. The right level depends on the task, the consequences of error, the reversibility of the action, the operating domain, and the system’s governance.
This distinction matters because capability, autonomy, and authority are different:
– Capability: what an agent can do – Autonomy: how independently it can proceed – Authority: where, when, how, and for whom it may act
A highly capable agent may properly remain at A1 when the cost of an error is severe. A simpler agent may operate safely at A4 in a narrow, controlled environment. As real-time human supervision declines, human responsibility remains. Governance must shift from the individual user to the agent, its operating environment, and the institutions around it.
Greater capability should not automatically bring greater authority.
Our governing principle is simple: No agent should operate at a higher level of autonomy than its governance environment can safely support. Transportation scaled because vehicles evolved alongside roads, standards, licensing, insurance, and law. AI agents will need their own rules of the road. The organizations that win will match each agent’s autonomy to its task, and prove their permissions, controls, and accountability can handle it.
In this week’s news, Google launched a universal Gemini agent for enterprise work, OpenAI introduced GPT-6 with interactive interfaces across ChatGPT, Google and Constellation agreed to expand nuclear power capacity by 890 megawatts, Anthropic launched a new AI-powered vulnerability scanner for open-source software, and Waymo secured $5 billion to accelerate autonomous vehicle expansion.
Full Weekend Edition below. 👇
Signals Shaping the Future of AI:
Infrastructure
Google signed a 20-year agreement with Constellation to add 890 megawatts of nuclear power capacity. The agreement covers upgrades to 11 reactors across six plants, alongside a separate 2,700-megawatt power purchase commitment. Click here
Broadcom, Oracle, and SpaceX are pursuing massive debt deals to finance AI chips. Broadcom is reportedly arranging more than $50 billion for OpenAI’s custom chips, while SpaceX is exploring $40 billion in financing for NVIDIA hardware. Click here
GlobalFoundries signed a $2 billion, five-year agreement with TSMC to manufacture silicon interposers for advanced AI chips. Production at its Malta, New York facility is expected to begin ramping in 2028, establishing a U.S. source for critical AI packaging components. Click here
CoreWeave announced plans to enter India with 240 megawatts of AI data center capacity through AdaniConneX. The expansion includes facilities in Navi Mumbai and plans to deploy NVIDIA Vera Rubin systems. Click here
Samsung projected a record $80 billion in third-quarter operating profit as AI memory demand surged. Preliminary earnings rose 782.5% year over year, driven by high-bandwidth memory demand and rising chip prices. Click here
Enterprise
Google launched Gemini at Work, a universal AI agent designed to execute work across enterprise applications. The agent connects organizational knowledge, business systems, and tools to plan and complete multi-step tasks with built-in governance controls. Click here
OpenAI launched GPT-6 with Intelligent UI across ChatGPT. The release enables ChatGPT to generate interactive interfaces, charts, and task-specific tools directly within conversations. Click here
Anthropic launched Claude Haiku 5.5, cutting average runtime costs by approximately 75% compared with Haiku 4.5. The faster model targets high-volume enterprise tasks, real-time assistance, and coding subagents. Click here
OpenAI and Atlassian expanded their partnership to bring GPT-6 models into enterprise workflows. OpenAI models will power agents across Atlassian’s platform and Rovo, combining frontier intelligence with organizational context from the Teamwork Graph. Click here
Capital Flows
Waymo secured $5 billion in its first debt financing to accelerate robotaxi expansion. Blackstone, PIMCO, Sixth Street, and other lenders participated as the company scales commercial autonomous driving. Click here
SAP agreed to acquire TechWolf to strengthen the enterprise context behind its Joule AI agents. TechWolf maps employee skills and work activities across business systems, helping SAP bring deeper workforce intelligence into SuccessFactors. Click here
AI chip startup Etched is fielding investment offers at valuations ranging from $40 billion to $50 billion. The early-stage discussions follow its $700 million financing at a $21 billion valuation in September, as demand for specialized inference chips accelerates. Click here
Arena raised $200 million at a $3.1 billion valuation to expand its AI model evaluation platform. Lightspeed Venture Partners and Khosla Ventures led the Series B as demand for model benchmarking and evaluation grows. Click here
Nous Research raised $90 million at a $1.5 billion valuation and launched Hermes for Businesses. The funding will support customizable AI agents that automate multi-step enterprise workflows. Click here
Research
OpenAI published a collection of mathematical results generated by an unreleased frontier AI model. The release includes hundreds of mathematical manuscripts and machine-checkable formalizations intended to support independent scrutiny and further discovery. Click here
Anthropic launched OSS Scanner to help open-source projects identify software vulnerabilities using advanced AI models. The company is offering eligible projects free recurring security scans, with reports generated directly by its models without human review. Click here
Anthropic used Claude Science to help produce the first complete ultraviolet map of the sky. The AI-assisted project reconstructed previously unmapped regions and included uncertainty estimates, demonstrating a new application of AI in scientific research. Click here
Policy
The White House established a new Super Intelligence Force to coordinate federal AI policy and assess emerging risks. The task force brings together officials from intelligence, defense, and regulatory agencies to examine AI safety, national security, and federal oversight. Click here
Rep. Lori Trahan introduced a discussion draft of the CLAIM Act to address liability for harms caused by autonomous AI agents. The proposal would establish legal accountability when AI systems engage in conduct that would be unlawful if performed by humans. Click here
U.S. lawmakers are advancing bipartisan legislation to change how data centers pay for power infrastructure. The proposal seeks to accelerate energy development while protecting residential customers from costs associated with large AI facilities. Click here
The FTC is advancing its investigation into frontier AI safety practices. The agency is reportedly preparing investigative demands as it examines how leading AI developers evaluate and manage risks associated with advanced models. Click here
Global AI Strategy
China’s AI sector is attracting major new investment as DeepSeek nears a $12 billion funding round and Manus raises more than $500 million. Backed by Tencent and other Chinese investors, the financings highlight efforts to retain domestic AI talent and strengthen China’s position in the global AI race. Click here
India is preparing a consultation paper on broader AI regulation within 30 days. The government is considering a technology-focused regulatory approach addressing AI risks while expanding national computing infrastructure. Click here
The UAE and U.S. are expanding cooperation on AI, cybersecurity, quantum computing, and advanced technology. A UAE delegation met senior U.S. officials and technology leaders in Washington to advance bilateral research and technology partnerships. Click here
The UK government accepted all 44 recommendations to overhaul AI regulation in healthcare. The new approach prioritizes continuous safety monitoring, clearer accountability, and updated approval pathways for AI-enabled medical technologies. Click here
Talent Signals
Each week, we spotlight key roles tied to the themes shaping this week’s AI headlines, connecting talent to the companies driving the news.
Opti is building an AI-native identity and access management platform designed to help enterprises manage access as users, applications, and AI systems interact across increasingly complex environments. The Mayfield-backed company is hiring across machine learning, software engineering, customer success, and enterprise roles. Click here
Goodfire is developing interpretability tools to understand, monitor, and steer advanced AI models as they become increasingly autonomous. The company is hiring across AI research, machine learning, engineering, security, and go-to-market roles. Click here
Mecka AI is building real-world data infrastructure for robotics and embodied AI, capturing human motion and video data to train physical AI systems. The company is hiring across robotics research, software, hardware, data, and deployment. Click here
You can see all the opportunities at Mayfield-backed AI companies here.
Social Signals
The most important conversations in AI are unfolding across social media, where top voices are shaping the next wave of signals and strategy. Here are some of the top social signals and their takes from the past week.
Ethan Mollick (Click here) — “It increasingly looks like we are going to have two revolutions from AI across academia: 1) Everything ever published will be re-read and re-judged in ways that human scholars never anticipated; 2) Novel discoveries will start to come faster than we can absorb.” Mollick argues that AI’s scientific impact could come from both revisiting the entire existing research corpus and generating novel work at a pace institutions cannot absorb. He points to OpenAI’s recent mathematics work and says nearly autonomous research in his own field is approaching top-journal quality.
Aaron Levie (Click here) — “The compute needed for the stage of AI we’re about to enter is going to be insane. Personal agents, agent swarms defending enterprises, agents that review all of our code for security issues, agents that process nearly all enterprise data in workflows, background agents working on tasks 24/7, and on and on.” Levie argues that the next infrastructure wave will be driven not only by bigger models, but by enormous volumes of persistent agent activity requiring compute, networking, file systems, and memory around the model itself. The post became one of the week’s more widely discussed infrastructure takes and was amplified by Elon Musk, who replied, “People have no idea.”
Matt Shumer (Click here) — “My budget is X. Find a way to track your token spend, and ensure you complete the task before the budget runs out.” Shumer argues that agents can already be given explicit economic constraints and then adjust how they plan, verify, and execute work within those limits. The broader point is that as agents become more autonomous, they will increasingly need to manage not only tasks and tools, but also the cost of their own execution. That introduces a new layer of control for enterprises: budgets can become part of the agent’s operating environment, alongside permissions, memory, and runtime guardrails.
