💡 The AI Agent Spending Problem Nobody Can Turn Off

In partnership with

☕ Morning! 💡 Your Weekly 5-Minutes of Caffeine and Tech Clarity

Quick Hits 🎯

• 🍎 One in five enterprises can't stop a runaway AI agent's spending in real time

• 💥 GitHub's August 17 outage just became the industry's newest case study in incident response

• 🔧 A new inference startup wants to sell tokens the way AWS sells compute, and it changes how you'll buy AI

• 🧭 Sequoia's latest playbook says owning your model and data stack now beats renting someone else's

• 🏢 State-backed hackers are quietly probing the infrastructure behind hospitals and power grids

• 🤖 One Slack message can now spin up a standing team of AI agents that remembers everything

• 💰 An AI-native accounting startup hit a $1B valuation two years after leaving stealth

• 🎰 Nvidia, OpenAI, and Google are all buying their way into each other's data pipelines

🎁 + 2 other stories you might find useful

Our Partner 🎉 

Your prompts are leaving out 80% of what you're thinking.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

The Big Picture 🖼️

💡 Nobody Can Turn Off Your AI Agents Fast Enough.

A fifth of enterprises running AI agents have no real-time way to shut off spend once an agent starts running wild.

That's not a hypothetical, agents booking compute, calling paid APIs, and spinning up resources are already outpacing the budget guardrails built for human employees.

The gap isn't the AI. It's the plumbing. Most finance and security tooling was built to catch a person clicking "approve," not a system making thousands of decisions a minute.

The data on where the spend is actually leaking shows most teams find out after the invoice, not during the incident.

The takeaway: if your org is shipping agents faster than it's shipping kill switches, you already have the problem, you just haven't seen the bill yet.

💡 A Major Outage Just Rewrote the Incident Response Playbook.

GitHub went down hard on August 17, and the postmortem that followed wasn't the usual PR softening. It named the failure mode, the blast radius, and the specific fixes going into the pipeline.

That matters beyond GitHub. Every team building on top of shared infrastructure inherits this risk, and the teams that recover fastest from outages, not the ones that never have them are the ones customers actually trust long term.

The full breakdown of what broke and what's changing is a rare look at a major platform showing its work.

The takeaway: reliability isn't measured by your uptime number. It's measured by how fast and how honestly you talk about the time you didn't have it.

💡 AI's Next Price War Isn't About Smarter Models. It's About Cheaper Tokens.

As agents take on more of the actual work, the cost that matters most stops being "which model scores highest" and starts being tokens per dollar and per watt at scale.

That reframes the whole stack. Companies optimizing hardware, caching, and scheduling together are pulling ahead of anyone still buying inference off the shelf. One inference startup's bet on owning that whole pipeline reads like a preview of where every serious AI infrastructure budget is headed next.

The takeaway: the model war gets the headlines, but the margin war is being fought in the inference layer, and most teams haven't started paying attention yet.

💡 Renting Your AI Stack Is Becoming the Expensive Option.

For the last two years, "just plug in an API" was the fast path to shipping AI features. That math is starting to flip.

Teams that own or tightly control their models, data pipelines, and orchestration move faster, hit fewer compliance walls, and aren't stuck waiting on a vendor's roadmap. Sequoia's field guide to making that shift lays out exactly where the line is between "fine to rent" and "you need to own this."

The takeaway: the companies compounding an advantage right now aren't the ones with the best prompt. They're the ones who stopped depending on someone else's.

💡 The Next Cyberattack on Your Company Might Not Be Aimed at Your Company.

State-backed groups are increasingly targeting the shared infrastructure underneath everyday systems, power, water, hospitals, not to steal data, but to have leverage ready if it's ever needed.

For anyone picking security tooling or a managed detection partner, this changes the calculus. You're no longer just defending against a competitor's script kiddie or a ransomware crew, you're one hop away from geopolitics. Just how exposed critical infrastructure really is is a lot more sobering than most security briefings let on.

The takeaway: "we're too small to be a target" stopped being true the moment your vendor's vendor got connected to the grid.

💡 Your Next Coworker Might Be a Slack Message.

A new integration lets anyone spin up a persistent team of AI agents from a single Slack message, agents that stick around, remember context, and keep working across conversations instead of resetting every time.

This is the consumer-grade convenience that enterprise tools have been slow to ship, and it's going to put pressure on every incumbent collaboration platform to match it, fast, and with the governance layer enterprises will demand. How the setup actually works is worth five minutes if your team lives in Slack.

The takeaway: the bar for "AI teammate" just dropped from a custom build to a single message. Expect your Slack channels to get more crowded.

💡 A $1B Accounting Startup Says Something About Where AI Budgets Are Actually Going.

Rillet just raised $100M at a $1B valuation, two years after coming out of stealth - in accounting software, a category most people would call boring.

That's the signal. AI-native workflows are winning in the categories everyone assumed were already solved, because "boring" software run manually is exactly where AI removes the most grunt work. The story behind the raise shows what happens when a team rebuilds a category from the data layer up instead of bolting AI onto the old workflow.

The takeaway: the next unicorn in your industry probably isn't building something new. It's rebuilding something you stopped thinking about.

💡 The Real AI Race Is Over Who Controls the Data, Not the Model.

Nvidia is backing OpenAI's data center buildout. Anthropic is making its own infrastructure moves. Google is buying up data assets in places nobody expected. None of this is really about who has the best model this quarter.

It's about who controls the pipes, the compute, and the raw data feeding the next generation of models because that's the layer that's hardest to replace once you're locked in. The full map of who's aligning with whom is the clearest read yet on where the real leverage in AI is consolidating.

The takeaway: watch the infrastructure deals, not the benchmark charts. That's where next year's winners are being decided right now.

Trending Tools 📈

PostHog/posthog (+14 ⭐ per day, 🐍 Python) Link

• All-in-one product analytics, session replay, feature flags, and experiments in one platform

• Helps teams: Cuts the number of separate tools needed to ship, measure, and fix product changes

makeplane/plane (+29 ⭐ per day, 🔷 TypeScript) Link

• Open-source alternative to Jira, Linear, and ClickUp for tasks, sprints, and docs

• Helps teams: Consolidates project tracking into one customizable workflow instead of stitching tools together

donnemartin/system-design-primer (+33 ⭐ per day, 🐍 Python) Link

• A structured guide to large-scale system design, complete with flashcards

• Helps teams: Standard reference for interview prep and leveling up engineers on distributed systems

langchain-ai/langchain (+12 ⭐ per day, 🐍 Python) Link

• Framework for building AI agents and multi-step LLM workflows

• Helps teams: Speeds up building autonomous agents without reinventing orchestration from scratch

microsoft/playwright (+10 ⭐ per day, 🔷 TypeScript) Link

• Cross-browser testing and automation for Chromium, Firefox, and WebKit

• Helps teams: One API for end-to-end tests across every major browser, cutting flaky test maintenance

nautechsystems/nautilus_trader (+53 ⭐ per day, 🦀 Rust) Link

• Production-grade, event-driven trading engine built for deterministic execution

• Helps teams: A serious foundation for building or backtesting algorithmic trading systems

pola-rs/polars (+2 ⭐ per day, 🦀 Rust) Link

• High-performance dataframe library built in Rust

• Helps teams: Handles large analytics workloads faster and with a lighter memory footprint than pandas

apache/kafka (+3 ⭐ per day, ☕ Java) Link

• Distributed event streaming platform for real-time data pipelines

• Helps teams: Backbone for event-driven architectures and real-time analytics at scale

Tech Trend of The Week 📊 

🔍 "AI data center" search interest is spiking, but not the way builders want

The searches climbing this week aren't about jobs or investment. They're about noise, water use, and rising electric bills near new sites, showing up in local ads and even campaign messaging.

The signal: data centers went from an abstract line item to a neighborhood issue in about one news cycle. Anyone planning a build should expect community pushback to be part of the timeline now, not a footnote.

Our Partner 🎉 

The best voice models, now with full orchestration. Build real-time voice and chat agents on one low-latency stack: any LLM, your tools and knowledge, testing, Guardrails, and omnichannel deployment.

Start building free

Hit Reply And Tell Me 💬 

What's the most impressive tech you've seen recently that actually works?

I read every single reply

Go where your customers actually are 📍

Most of the internet copies Reddit.

Blogs, SEO pages, AI models, product reviews. They all pull from the same well.

Reddit is the 3rd biggest site on earth. And it trains the bots your users ask for help.

Odd Angles Media runs Reddit campaigns for over 45 brands each month.

Luckily, we’ve convinced them to give away all of their strategies in a (free) ebook ⬇️⬇️

Grab The Free Ebook (Free) →

Reach the People Who Sign the Checks 💰️ 

40,000+ CTOs and engineering leaders. 96% US-based. 80%+ corporate emails.

They evaluate vendors. They shortlist solutions. They buy.

Advertise with us on Paved → (discount applied)