Welcome, humans.
Okay, so I heard an interesting prediction this weekend: who gets blamed when AI agents do something bad is the #1 open question of the agent economy.
On More or Less, the group says AI safety gets weirder when we consider our new “agent ecology”: potentially billions of agents researching, negotiating, buying things, writing software, and dealing with other agents on our behalf.
When this happens, alignment gets messy. Dave Morin’s question: does your agent actually represent you, or the company that built it?
Sam Lessin thinks liability law may decide it.
If the AI company is liable, he predicts it will optimize for self-protection.
If users are liable, incentives shift toward user control.
Across thousands of services, whoever bears responsibility could shape the ecosystem.
Now, that’s a prediction, not the current legal framework. The point: companies will ship agents before society decides who pays when…
Hacks something…
Makes a bad purchase…
Breaks another service…
Or causes damage.
The panel says frontier labs could face losses far beyond normal software failures, what Sam calls “infinite liability,” unless governments absorb some risk.
So we may spend the next few years obsessing over agent benchmarks, only to discover that one of the most important AI alignment mechanisms is… tort law.
Here’s what else happened in AI today:
😺 Meta, Manus and Instinct raced for your main agent.
📰 Florida sought emergency restrictions on OpenAI models.
📰 Researchers warned AI could accelerate AI research.
🍪 NVIDIA moved agent safety below the model.
🎓 Jev routed cheap decisions away from big models.
ICYMI: our recent Deep Dive covers an AI agent that created fake identities to pressure a real person, plus the common company risk: ordinary tools with access nobody is watching. Special shout out to Harmonic Security for sponsoring this one!
😺 Meta, Manus and Instinct Are Racing to Become Your Main AI Agent…
Yesterday, there was a flood of agent announcements that looked like a land grab for the layer between you and the internet, dressed up as adorable little assistants!
Meta finally has its Muse; Manus launched Manus 2.0 and spun-up Cue; and viral invite-only agent Instinct became a Silicon Valley darling with a $1B Series C at a $10B valuation, one month after $350M at $2.5B.
What’s going on here? Different layers, same direction: AI is moving from something you open to something that represents you* (well, per above does it??), with a computer, memory, credentials, permissions, and autonomy.
Examples:
Muse gets a secure VM and browser.
Manus 2.0 gets persistent cloud computers and automations that keep running after you leave.
Cue gets a phone number, email, wallet, and computer.
The cuteness masks the business model. Ben Thompson’s Stratechery frames these agents as the next, most ultimate aggregator: Publishing abundance let Google and Facebook aggregate discovery; action abundance lets agents aggregate where you shop, book, compare, negotiate, build, research, and pay.
So what does this mean? Whoever owns the interface sees your intent first, and aggregates everything else.
Want a Chicago flight under $500? Your agent picks the service, tradeoffs, and supplier. The airline supplies the seat; the agent owns the relationship end to end.
Dave Morin of More or Less sees it differently. He says the future will be more “polyagentmorous”: every app becomes an agent, with “no one agent to rule them all.” What connects them? “THE INTERNET.” Or should we say… Metaverse?
The new Prime Agent previews that architecture: persistent sub-agents that message, preserve memory, and update their setup. Using Claude Cowork or ChatGPT Codex (or other agents) gets you used to it, too: calling up subagents to do work for you from your main agent chat.
This pattern of work is used more for coding and research today, but the model is the same: your main agent boo delegates to your side piece agents.
They’re just there to get the job done, you don’t care about maintaining the longevity of the relationship, ya know? How’s that for polyagentmory, Dave!
Why this matters: If you run a business, becoming everyone’s main agent means fighting Meta, Microsoft, OpenAI, Anthropic, Google, Manus, Instinct, and the open agent ecosystem. You don’t want that smoke, bro.
Better: make your goods and services easy for somebody else’s agent to use.
Expose machine-readable inventory, availability, pricing, and policies.
Offer reliable APIs, MCP connectors, or structured actions.
Make authentication, permissions, purchases, and refunds agent-friendly.
Be the supplier the aggregator picks.
Computer-use agents can click almost any website, so “we don’t have an API” isn’t much of a moat. It makes you slower and error-prone.
Muse’s momentum matters because Meta combines distribution, a dedicated computer, payments, connectors, context, and an enterprise extension… with Meta’s instant distribution to 3.6B daily active users behind it.
The internet was built around getting humans to click your website. The next one may be built around getting an agent to pick you.
FROM OUR PARTNERS
Ready to move from AI experimentation to execution?
Choosing the right model is only the beginning. Is your data ready? Where should you deploy AI? How will you measure ROI?
MIT xPRO’s Deploying AI for Strategic Impact helps professionals answer these questions and turn AI initiatives into business value.
And MIT xPRO learners are seeing an impact: 42% reported gaining new responsibilities, earning a promotion, or moving into a new role after completing the course.
Learn from MIT faculty and industry experts and build a practical framework for deploying AI strategically.
🎓 AI Skill of the Day: Branch a good ChatGPT thread instead of starting over
Ever get 20 messages deep and want another direction without wrecking the thread?
On ChatGPT web, you can “branch” from any message. The new chat keeps everything before it; the original stays untouched.
Hover over the message.
Click More actions (⋯), then Branch in new chat.
Try the alternate plan, rewrite, or debugging path.
Basically Git branches if you’re technical, but for the conversation you were afraid to touch. OpenAI says it works for logged-in web users, including Projects. Try it!
Have a specific skill you want to learn? Request it here.
FROM OUR PARTNERS
Trigger.dev’s chat.agent gives every conversation its own durable machine, so the API route you’d normally maintain just disappears. Build with the AI SDK you already use; turns run without serverless timeouts, survive tab closes, and pause at zero idle cost.
🍪 Treats to Try
*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!
*See how Adobe is bringing AI to enterprise documents to help teams find insights and work faster. Read More.
Claude Sonnet 5.5 is Anthropic’s faster Sonnet update, using fewer tokens with strong coding and agent performance; read the prompting guide here, plus the distilled top 12 tips to prompt Opus.
Meta launched its Enterprise Platform, bringing Muse, Muse API, Muse Code, and its AI stack to work.
NVIDIA launched its Open Agent Safety Platform, pairing sandboxing with a hardware watchdog that enforces rules outside the model.
Perplexity Agent API lets teams define reusable agents with versioned Profiles, Skills, and managed connectors.
Mo is Momentic’s scriptless AI QA engineer: tell it what to test in plain English, and it explores your app, confirms bugs, then returns repro steps, logs, and video.
Anthropic’s claude-api eval + hillclimb workflow helps you prove an agent change actually made things better, not just different: it tests changes against realistic tasks and unseen examples, then keeps improvements and rolls back changes that only looked good on the practice set.
Eleven v4 handles expressive speech across 90+ languages; Turbo adds real-time use, performance direction, and multi-speaker dialogue.
📰 Around the Horn
OpenAI reportedly scrapped the release of GPT-6.1 Astra for tomorrow’s DevDay after internal safety tests found more deception and scope-authorization failures.
AMD agreed to acquire Fei-Fei Li’s World Labs for about $8.2B in stock, adding world models for physically grounded 3D AI.
Anthropic’s IPO prospectus leaked, with key highlights including: a valuation above $2T after $4.6B in 2025 revenue alongside an $8.06B operating loss; nearly 25% concentration of customers across two clients (!!); and roughly $518B in pledge future compute and infrastructure commitments.
Florida’s attorney general asked a state court for an emergency injunction restricting OpenAI from developing new models without independent safety safeguards.
President Trump, Speaker Mike Johnson, and AI executives are expected to meet today at the White House to discuss AI risks and policy.
A Cambridge-led report warned automating AI R&D could compress years of progress into months and urged governments to measure it.
🧰 Tuesday Tool Tip: Stop paying a giant model to make tiny decisions
AI workflows waste money asking huge models to write when all you need is a decision.
As Victor Dibia writes, Jev skips prose: it scores predefined choices and returns the winner with confidence. Think “sales / support / billing,” “yes / no,” or “cheap model / expensive model.”
That changes the workflow: use expensive generative models only when generation is required.
List recurring decisions with a small answer set.
Route them through a decision model or classifier.
Escalate uncertain or open-ended cases to your frontier model.
A useful rule: if you can write every valid answer on a sticky note, you probably don’t need a giant model composing prose to pick one.
See Jev’s probability scoring and calibration in Victor Dibia’s technical explainer.
New from The Neuron: AI Explained
Click the image above to learn how data centers actually work!
A Cat’s Commentary
Enlightened? Wow! High Praise! Call us the John Locke and Voltaire of AI I guess!
 | That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today!
|
