Who Owns the Agent Era? Anthropic Courts Founders, Temporal Buys Security, Zscaler Bets Its SOC

LaurenOkay, so get this. Everybody and their cousin is selling an AI agent right now. Your inbox has one. Your calendar has one. I think my toaster is about to pitch me one.

DerekAnd every single one of them needs a place to live.

LaurenWhich brings us to Tech Insider Weekly. I'm Lauren.

DerekAnd I'm Derek. Today's big question: Who gets to keep the keys?

LaurenThe stuff around the agent, to be clear, the layers underneath and on top that make it usable.

DerekOh, the boring plumbing.

LaurenThe boring plumbing where the money tends to land.

DerekWait, so are we saying scrappy startups build all of it, and then somebody huge just walks in?

LaurenPlot twist, maybe. Or maybe the startups hold on. We've got three stories that pull on that question from different directions.

DerekOkay, so which story do we start with?

LaurenOh, you're gonna love this one. First, the company that wants to be the home base for founders building with AI.

DerekA landlord for the AI era. I'm already suspicious.

LaurenWait for it. Okay, so get this. October 6th, Anthropic announces it's expanding its Claude Startups program, and the pitch is basically a welcome basket for founders.

DerekA basket with what in it?

LaurenAnthropic's own announcement says membership includes a year of Claude Team plan, API credits, special offers on tools for running an AI native company, and applied AI office hours.

DerekOffice hours, so a human at the company will actually talk to you.

LaurenApparently, yes.

DerekThat's the part I'd take, honestly. Credits run out. Someone who's seen a hundred agent builds before you can save you a month.

LaurenThough I'd wanna know how much time they actually give you. The announcement just says office hours, and I can't tell you what the format looks like.

DerekRight. So treat it as a bonus until you've tried it. For a tiny team, even one good conversation about how to structure an agent could be worth more than another discount code. That's my opinion, not a promise from anybody.

LaurenFair, but the numbers are where it gets silly. CNBC reported up to $45,000 in Claude Startup Stack discounts and credits.

Derek45 grand?

LaurenUp to, plus a one-time thousand dollar API credit.

DerekHa. So the headline number is the buffet, and the API credit is the free sample.

LaurenRude, but yes. And CNBC says the free year of Claude Team covers up to five seats.

DerekFive seats is a whole early startup. Who qualifies?

LaurenCNBC says startups founded within five years or funded within the last two can apply.

DerekWait, five years old counts as a startup? In this economy, that's practically a retiree.

LaurenRight? That's a wide door.

DerekAnd notice it's two ways in: young company or recently funded company. Either clock can get you through.

LaurenWhich means a founder listening on a lunch break should check both clocks before assuming they're out. You might be older than you think and still qualify, or the other way around.

DerekAnd the smart move is to read the actual terms before you plan around any of those numbers. Up to forty-five grand is a ceiling, not a guarantee, and I wouldn't budget a runway around a ceiling.

LaurenYes. Treat the big number as the best case and the thousand dollar credit as the thing you can count on.

DerekSo why open it wide now?

LaurenCNBC frames it as Anthropic building its founder base ahead of an expected IPO. That's CNBC's read, to be clear, not something Anthropic said.

DerekAnd a Briefs piece describes the same moment. Lotta people expecting a blockbuster IPO.

LaurenWhich is my skeptical brain talking now. Free credits are a customer acquisition move. Dressed up nicely, sure, but it's customer acquisition.

DerekThough to be fair, customer acquisition isn't a dirty phrase. A company wants founders to like its product enough to build on it. The real question is how sticky the free stuff makes it.

LaurenRight. Free isn't evil. It's just never free, free. There's a business reason behind the basket, and it's fine to say so out loud.

DerekOh, totally. But I'd go one step further. Credits shape what founders build on.

LaurenMeaning?

DerekYou get a year free, you wire your product to one model's quirks. Your team learns its habits. Month 13, switching sounds like a rewrite.

LaurenOkay, that's the sneaky part. The bill isn't the lock-in, the muscle memory is.

DerekIn my opinion, anyway. Maybe founders hop around happily.

LaurenMaybe. I'm not betting my savings on it.

DerekHere's my practical advice, though. Take the perks, but keep your model calls behind a thin layer of your own code so switching stays possible if you ever want it. Then the year free is a gift instead of a gravity well.

LaurenSo the founder version of dating without moving in together. Love it.

DerekAnd meanwhile, all those founders are about to build agents that go off and do things.

LaurenOh, you're gonna love where this goes. Once founders build on agents, somebody has to stop those agents from doing things they shouldn't.

DerekWho's even on that job?

LaurenOkay, so get this. Agents need a babysitter, and a company in Bellevue, Washington, just hired one. Temporal is bringing over a team from a New York startup called Oso.

DerekThe Oso folks. Security controls for AI agents, right?

LaurenRight. GeekWire reported it on October seventh. Temporal hires the team. Oso's people go build the guardrails inside Temporal.

DerekWait, so why does an agent even need guardrails? People manage fine.

LaurenDo we, though?

DerekFair, but Temporal's own blog post has an actual answer. It says agents lack the restraint people have.

LaurenRestraint, like a coworker who sees the admin password and thinks, "Huh, not mine."

DerekExactly that. A person hesitates. An agent just runs. So Temporal's argument is that authorization has to live inside the execution platform itself, not get bolted on afterward.

LaurenHold on. Inside the platform, meaning the thing running the agent is also the thing saying no?

DerekYes. Same place, same moment.

LaurenOkay, that's neat. But give me the scary number. There's a scary number. I can feel it.

DerekThere is. Temporal's blog cites Oso research saying ninety-six percent of permissions go unused.

LaurenNinety-six. That's like handing out a hundred keys, and ninety-six of them never touch a lock.

DerekAnd note where it comes from. That's Temporal's blog citing Oso. Nobody independent has checked it that we know of, so hold it loosely.

LaurenNoted. Still a juicy number, though.

DerekTotally. Let me make it plain. Say you give an agent a house key. It opens every door in the place, but the agent only needs to water the plants in one room.

LaurenAnd instead, it can wander into the bedroom, the safe, the garage.

DerekNobody's saying it will. The point is that it could, and nothing stops it.

LaurenHmm. Okay, but it's one run, right? One job?

DerekThat's the other piece. Temporal says a single agent run can chain many calls across services. Call one hits a database, call two hits a payments tool, call three hits email.

LaurenSo the whole chain rides on one key.

DerekWhich is why Temporal says the identity behind each call matters. Every hop needs to know who is actually asking.

LaurenOoh, so it's not one front door check. It's a check at every door in the hallway.

DerekEvery single one.

LaurenOkay, Derek, translate for a founder, somebody with six people and an agent demo. What changes for them on Monday?

DerekHonestly, the first change is a question. When you wire your agent to your tools, what did you hand it? Most teams hand over the master key because it's the fastest way to get the demo working.

LaurenGuilty. I've shipped that exact shortcut and told myself I'd fix it later.

DerekEverybody does. Later is when the agent does something weird at three AM.

LaurenAnd you find out what Ninety-six percent unused really means.

DerekRight. So the practical shift is scoping. Give it the one room and have the platform check every call so you're not writing that logic yourself.

LaurenWhich is the appeal of leaning on a platform instead of building it all yourself, and you can list your agent's tools and write down what each one actually needs. That's an afternoon of homework, and it shows you the gap between what you granted and what you use.

DerekAnd that's the bet Temporal is making. It hired the people who think about this all day.

LaurenCareful, though. What I've got is that Temporal hired the team. I can't tell you any terms of the arrangement, so we shouldn't guess at them.

DerekFair. My read, and it's only a read, is that the expertise is moving inside the platform instead of living as a separate tool you bolt on.

LaurenAnd what we can say is that the people who worked on permissions are now building them where agents actually run.

DerekWhich is the direction that matters. Less stitching together, more of it built in.

LaurenSo a startup team is joining a platform. Switching gears, though, because a big incumbent has just launched its own take on the same problem.

DerekOh, you're gonna love this one. The established names aren't sitting this out.

LaurenOkay, so get this. While Temporal is out hiring a startup team, an incumbent walked into the same neighborhood with its own product.

DerekZscaler.

LaurenZscaler, and they called it Agentic SOC. Derek, you have to translate SOC for the people in the car.

DerekSecurity Operations Center, the room, real or virtual, where people stare at alerts all day and decide which ones are fires and which ones are a raccoon in the server closet.

LaurenMostly raccoons.

DerekMostly raccoons, and that's the pain here. Too many alerts, too few humans.

LaurenSo Forkast ran a piece on October seventh. The framing is that a thirty-three billion dollar incumbent is betting its SOC on AI agents.

DerekThirty-three billion. That's not a scrappy team hoping for a lucky quarter.

LaurenRight. And Forkast says it raises the question of whether startups or platforms end up owning the security layer.

DerekWhich, hello, is our opening question wearing a security badge.

LaurenHa. Okay, but what does the thing actually do?

DerekSecurity trade coverage from SecurityBrief says it pairs Zscaler's own telemetry with models from Anthropic and OpenAI. The telemetry number is seven hundred and fifty billion zero trust transactions a day.

LaurenA day?

DerekA day. That's the pile the agents get to read, and it's a staggering amount of traffic for software to sift through.

LaurenOkay, that's the real flex. The models are rentable. The pile isn't.

DerekThat's my read too. Anyone can call an API. Not everyone sits on that much traffic.

LaurenBut does a big pile automatically mean better verdicts? Data's only useful if the agent reads it well.

DerekRight, and that's the open question. More traffic gives the agents more context to work with, but nothing I have says how well they use it. Call the pile an advantage on paper and let the results speak later.

LaurenOh, wait, and both Anthropic and OpenAI? Not picking a side?

DerekThat's what the coverage says, models from both, which fits the earlier conversation, honestly. The model is the interchangeable part.

LaurenPlot twist for the founders with the free Claude year, I guess.

DerekGently, Lauren.

LaurenFine. So the agents do what exactly?

DerekPer the same coverage, triage, root cause investigation, verdicts, and response. So not just flagging a thing, deciding what it was, and acting on it.

LaurenHold on. Acting. Agents that act in a security room after we just spent a whole segment on agents needing limits.

DerekYes, which is why I want to separate what's shipped from what's promised.

LaurenOkay, what's promised?

DerekZscaler's own annual filing with the SEC, the 10-K for fiscal twenty twenty-six, describes AI agents that triage, investigate, and respond at machine speed, and it says threats can be contained before they reach the endpoint.

LaurenMachine speed before they reach the endpoint. That's a sentence with a lot of confidence in it.

DerekIt is, and those are the company's own claims. A filing is the primary document, so it tells you what Zscaler says the product is. It doesn't tell you how it performs in a messy customer environment.

LaurenFair. Nobody's published the real results.

DerekNot yet.

LaurenSo here's what I find funny about the timing. Zscaler launched this recently. Temporal's Oso hire is this week's news.

DerekSame question, two answers arriving close together.

LaurenExactly. One side, a startup team joins a platform. Other side, a big incumbent ships its own agentic product in the category.

DerekAnd I don't think either one settles it. Temporal brought in people who know authorization. Zscaler's leaning on the data it already has.

LaurenOkay, but if you're a startup security firm, which headline scares you more?

DerekHonestly, in my opinion, the Zscaler one. Getting out-distributed by someone with that much traffic is a way down.

LaurenOoh, but the counter, and again, that's opinion, is that a filing is a company describing its own product, so we should wait for results.

DerekSure. Could go either way. A startup can be faster and sharper. A platform can be everywhere already.

LaurenSo nobody wins on the day of launch.

DerekNobody wins on launch day. They win on whether the thing holds up at 3:00 in the morning with the raccoons.

LaurenAnd honestly, I like that this is still a debate. If someone tells you it's already decided, they're selling something.

DerekOr they own the pile. Either way, I'd rather watch what customers say after a few quarters than what the press releases say.

LaurenOkay, so what should a founder do with all this? Last stop. My opinion, and it's only that, pick a layer and be specific about it. Not AI security. One narrow job that one kind of customer hates doing by hand.

DerekNarrow like what?

LaurenNarrow enough that a platform's product manager would have to squint to see it. Big players ship wide. Wide is where they live.

DerekHm. I buy that. But narrow can also mean small, and small gets stepped on.

LaurenSure, hence the specific part. Specific people, specific pain.

DerekGive me a scenario, though.

LaurenSay a two-person team picks one chore that a security team does by hand and does that really well. A platform may not prioritize it, but a customer will happily pay to stop doing it. Again, just my opinion.

DerekFine. My opinion, also just an opinion: build the trust and the controls in from day one. Don't bolt them on.

LaurenSay more.

DerekBecause the buyer who lets an agent touch their systems is gonna ask who approved what and what it was allowed to do. If your answer is, "We're adding that next quarter," the deal's already dead.

LaurenThe roadmap slide of doom.

DerekEvery startup has one. It always says Q3.

LaurenPicture a founder in a security review with no record of what the agent did. They're reconstructing it from guesses while the buyer watches.

DerekThat's a bad afternoon. The version where you can show who approved what is a completely different meeting. Opinion again, but I'd rather be the team with the boring answer ready.

LaurenWait, though, doesn't that slow you down? Controls are the boring stuff.

DerekSlower on week one, sure. Faster on the day a security team reviews you. Plus, it's the part an incumbent can't easily fake because they're retrofitting.

LaurenOoh, okay, that's a good one.

DerekThank you. I'll be insufferable about it.

LaurenYou already are. One last angle from Forkast, whose headline says Zscaler's launch signals the category startups built is now an enterprise requirement.

DerekWhich is a strange compliment, right? Your whole category got validated by the guy who might eat you.

LaurenExactly, and the practical meaning is that a buyer's going to expect it. Nobody gets to treat agent security as a cute experiment anymore.

DerekWhich loops back to your pick a layer thing. If it's a requirement, there's a checklist, and checklists have gaps.

LaurenGaps are where small teams eat.

DerekMm, or get eaten. Depends on who builds the controls first.

LaurenWow, comforting. Anyway, that's the week. Let's wrap up. Okay, so Anthropic hands founders credits, Temporal hires the Oso team, and Zscaler bets its SOC on agents. Three stories, one question about who holds the keys.

DerekAnd nobody gets to say they've settled it.

LaurenNot yet. My takeaway for founders: watch which layer a big company decides to ship next. That's your early warning, and it's free.

DerekFree is my favorite price.

LaurenMine too. Derek, you're up.

DerekTechnical lesson. Decide what an agent is allowed to do before it does it. Build that check-in on day one, because bolting it on later means rewiring everything the agent already touches.

LaurenMm, retrofits are miserable.

DerekAsk anyone who's done one.

LaurenAnd if you've already shipped the shortcut, no shame. Just schedule the cleanup before the agent schedules it for you.

DerekThat's the whole lesson in one sentence, Honestly. Small fixes now, fewer surprises later.

LaurenOkay, housekeeping. Subscribe wherever you listen, so new episodes just show up.

DerekLeave a review too. It helps other people find the show, and we read them.

LaurenAnd send us ideas, Founders we should talk to, topics we should chase. If you're building in this space, we want to hear what you're running into.

DerekNew episodes land every Wednesday.

LaurenUntil then, go check your Agents' permissions. Seriously, right now.

DerekBefore coffee even.

LaurenThanks for listening to Tech Insider Weekly, and we'll see you next Wednesday.