AI's biggest questions just got real.
This year an AI broke an 80-year-old maths conjecture. Two months later, AI agents broke out of a lab test.
I worked with Claude to create a short video: what to be excited about, what to be wary of, and our calls on 7 big questions for 2027 and beyond.
We think by late 2027, AI will do most of the research behind the next AI. Do you agree?
Sources:
Capital One Ventures, Capital One's corporate venture capital arm (23 Sep 2025): https://t.co/mk3tKGnP7C
TechCrunch on Palo Alto Networks and Clutch Security (5 Nov 2026): https://t.co/XINYaOQP3l
TechCrunch on what Clutch secures: API keys, tokens and other credentials (29 Jan 2025): https://t.co/cY5aYsYomc
SiliconANGLE on how Bay checks agent actions (10 Nov 2026): https://t.co/TUyMVn1AJw
SecurityWeek on WhiteMatter's Series A (10 Nov 2026): https://t.co/ZXagGHBEAD
Sentra's announcement, PR Newswire (4 Nov 2026): https://t.co/4jMtT0Ba6N
Bay's announcement and its founders' coding agent, Business Wire (10 Nov 2026): https://t.co/xOh1HmeeGN
AI agents need four controls that banks already run on staff.
An agent that reads files, logs into systems and takes actions has the reach of an employee. Between 4 and 10 Nov 2026, a startup building each control raised money or was reportedly in takeover talks.
1. Entitlements: who may see and do what. @sentra_security raised $150M on 4 Nov in a round led by the venture arm of the US bank Capital One. Staff get access by role. Sentra starts from the data itself and controls which people and agents can reach it.
2. Logins: every login belongs to one named person. On 5 Nov TechCrunch reported that Palo Alto Networks was in talks to buy Clutch Security for $500M in cash plus milestone payments. Many agents log in with keys and tokens issued to software instead of a person. Clutch secures those keys and tokens.
3. Maker-checker: a second person approves an action before it goes through. Bay launched on 10 Nov with $50M from two rounds. @BatteryVentures led the $37M Series A. Bay's software blocks agent actions it judges too risky and sometimes asks the user to approve a step.
4. Supervision: someone reviews the work. WhiteMatter raised a $20M Series A led by @BlumbergCapital on 10 Nov. Here the supervisor is itself an agent. WhiteMatter says it can assign a dedicated guardian agent to every agent.
The checker matters most. Bay says its founders watched one of their own coding agents try to expose a cloud environment to the internet to complete a test. The agent had not been hacked. It "used legitimate permissions to pursue a legitimate goal in a dangerous way." Tighter entitlements help but are set in advance. Only a check before the action runs judges the step itself.
My take: whoever gives an agent its task is the maker. At a bank the maker never approves their own action. A second person on every step would lose the speed agents bring. Above a set limit, someone other than the agent's user should approve. The same goes for any step that opens up access.
Sources:
Anthropic, Claude Fable 5.2 and Mythos 5.2 (9 Nov 2026): https://t.co/yDKB7AN5KF
Claude Fable 5.1 prices, Claude Platform docs (checked 9 Nov 2026): https://t.co/UJzE0u5ofw
Remote Labor Index, Scale AI and the Center for AI Safety (29 Oct 2025): https://t.co/yWKuE3CusE
Anthropic's Claude Fable 5.2 outscores 5.1 and costs 20% less per token.
Anthropic says Fable 5.2 completed 41.3% of the paid freelance projects in the Remote Labor Index. Fable 5.1 managed 17.9%.
My take: compare models on cost per finished job, not price per token.
@gaspardlezin Samsung put the reach at 82M compatible Galaxy devices in the US. Stablecoin supply only grew 3% in the year to October, so wallets like this are where new users come from now.
Sources:
Daytona (the European tech news site https://t.co/80lHEtfC12, 20 Oct 2026): https://t.co/zPyaXjgCse
Prime Intellect (the US tech news site TechCrunch, 8 Jul 2026): https://t.co/Ej4sMgGfRD
Firecrawl (company blog, 22 Sep 2026): https://t.co/VYtWNxy0uD
Temporal, led by Lightspeed with co-leads Wellington Management, Goldman Sachs Alternatives and Tiger Global (company news, 14 Sep 2026): https://t.co/cvGTr1xLbe
Natural (press release, 20 Jul 2026): https://t.co/iPkzeYDRSY
Cymphony (press release, 9 Sep 2026): https://t.co/hHagy4u4i6
Rein Security (the business news site Pulse 2.0, 8 Oct 2026): https://t.co/y1a6k1H6QZ
AIR (press release, 1 Sep 2026): https://t.co/0DpM4hxHx1
AIUC (press release via Tiger Brokers, 15 Sep 2026): https://t.co/UQvtziOrCn
Raised an agent infrastructure round since July 2026? Reply with the announcement.
AI agents are getting what a new hire gets on day one.
A computer, a company card, a keycard, a manager, even insurance. Since July 2026, Sequoia, Lightspeed, Ribbit and others have backed nine startups building that kit. Five make agents more useful, and four make them more trustworthy.
More useful:
→ A computer: Daytona @daytonaio runs cloud computers for agents. $60M, FirstMark
→ Training: Prime Intellect @PrimeIntellect lets companies train their own agents. $130M, Radical Ventures
→ Something to read: Firecrawl @firecrawl turns web pages into clean data. $75M, Smash Capital
→ A to-do list: Temporal @temporalio lets agents resume long tasks after a crash. $550M, Lightspeed
→ A company card: Natural @naturalpay lets agents pay and get paid. $30M, Kirsten Green at Forerunner
More trustworthy:
→ A keycard: Cymphony tracks what agents can access and do. $30M, Sequoia and SMBC
→ A manager: Rein Security watches agents and blocks risky actions. $25M, Glilot Capital and Sienna Venture Capital
→ Building security: AIR @air__security runs a firewall for agents. $50M, Sequoia and Greenoaks
→ Insurance: AIUC certifies and insures agents. $40M, Ribbit Capital
My take: agents are resembling employees more and more. They're starting to look like they can legitimately take on the work of new hires.
@MagnusGrimeland Local presence matters even more in fintech. So much of it comes down to knowing how the regulators and banks work in practice, and that's hard to learn from a distance.
@stabledash That's what I like about DTCC's tokenised shares. The real shares stay where DTCC already holds them, so banks don't have to hand custody to a new firm.
@minchoi Calls are where this gets really useful. Most of the annoying stuff sits with businesses that only answer the phone, like a dentist or a plumber.
@ivanburazin Makes sense that the sandbox becomes the lab. If an agent is driving the fine-tune, you want it somewhere it can make a mess and start over without touching anything else.
@sytaylor Liability feels like the real blocker. Until someone decides who pays for an agent's wrong purchase, a card locked to one merchant type and one amount is the easiest thing for a finance team to approve.
@stabledash The Android part is doing a lot of work here. An iPhone app can't read your texts or other apps' alerts, so this only works where most people carry Android.
@suraj_sharma14 This is the one that makes voice agents feel broken in a demo. The agent cuts itself off mid-sentence and everyone assumes the model got confused.
@gaspardlezin Lines up with Mastercard's Q3 call, where it said agentic commerce volumes aren't significant yet. Hard to justify bigger rounds before the volume shows up.
@iamlukethedev Locking a skill is the one I'd turn on first. If an agent can rewrite its own instructions, you only find out it drifted when the output changes.