Steve Sanderson

Norwegians call it “Helt Texas.” I call it business as usual in most growth-stage companies. Every growth-stage founder/ CEO hits the same wall. Fighting fires, working ad hoc, taking things as they come. It all comes to bite you in the butt as your company scales. Most growth-stage companies I know (50MNOK - 300MNOK) resist putting an operational model in place. They think they're staying scrappy and agile. But here's what actually happens as you grow: → Teams hold different views of what matters most  → Decisions get made in Slack threads and forgotten  → Good people leave because they don't know what's going on  → Customers experience the internal chaos as inconsistency This is what Norwegians would call "Helt Texas." (crazy) The best-run companies I've worked with all have the same thing in common. They use a simple system that keeps everyone aligned without constant ad-hoc coordination. The basic framework looks like this: • Weekly leadership meeting. Same day, same time. • Quarterly all-hands. 1x/ quarter • Regular team meetings that cascade information up and down. • Space for informal connection (yes, this matters too). It's not complicated, but you do have to do it. Everyone thinks that go-to-market means how you show up on the outside. I take a more holistic view - How you show up on the outside is a direct reflection of how you work on the inside. i.e. The company that operates in chaos will project chaos to the market. So how do you make that chaos a little more manageable? That's the topic of this week's newsletter: https://lnkd.in/eWNU2cFk

13

11 Comments

"When you're validating the work to be done, you're also validating who you should be hiring to do it." Amos Schwartzfarb, executive coach, former managing Director of Techstars Austin and author of Sell More Faster and Levers, talks about getting roles right held on Tech Scenes Unplugged. #venturebacked #vcbacked #venturecapital #leadership #ceo #founder #operatingsystem #peakOS #foundermode #techstars

7

2 Comments

Louisiana is a one-party state. Gerrymandered maps. Voter purges. Polling place closures. The people most hurt by the system are the ones most blocked from changing it. This is Jim Crow. Updated. Still running. The design hasn’t changed since 1890. Only the tools. Ten Commandments in every public school classroom. Anti-vaccine bills. A resolution supporting the elimination of the Department of Education. Prison population up 2,000. Parole abolished. $1.3 billion for incarceration — from the same budget that funds schools. An immigration detention camp at Angola. Cancer Alley: cancer risk eleven times higher than government estimates. Trump’s EPA gutted the one rule that would have cut that risk by 96 percent. Landry called climate change a hoax and put a fossil fuel executive in charge of energy policy. Natural gas reclassified as “green energy” using a Koch-funded template. A $2.4 billion toll bridge was handed to a private company for fifty years. Cassidy — a doctor — calling it “reasonable” to cut 158,000 people off Medicaid when 69 percent of them already work. Dr. Isolde Butler at CrescentCare in New Orleans sees what happens next. Patients lose coverage, skip their blood pressure pills, and wind up in the emergency room with heart problems that didn’t have to happen. We pay for it anyway, just in the worst possible way. More prisons. Fewer schools. Poisoned air. Sold-off infrastructure. That’s the one-party state. And then there’s the corruption. Follow the money, and it leads right back here. People Who Think, a Republican ad firm in Louisiana, got $77 million in no-bid DHS contracts for Kristi Noem’s $220 million vanity ad campaign — the one where she sits on a horse at Mount Rushmore telling immigrants to leave. The firm’s head worked on Jeff Landry’s gubernatorial campaign with Noem’s chief advisor. The contracts bypassed competitive bidding by invoking a “national emergency.” Even Republican Senator John Kennedy called it out — and then, as usual, did nothing about it. That’s how it works. They name the corruption. They participate in it. Same party. Same money. Same silence when it counts. GEO Group — the largest private prison operator in America — reported a 700 percent profit increase last year. $254 million. Record-breaking. Their CEO called it “the largest amount of new business we have won in a single year in our company’s history.” In Louisiana, seven of nine ICE detention facilities are former state prisons and jails, repurposed to hold immigrants. Detainees work for a dollar a day. GEO charges them $11 for a tube of toothpaste. A federal judge found the company had enjoyed “unjust enrichment” through illegal labor exploitation. The Big Beautiful Bill gave ICE $45 billion to expand detention. Private prison companies donated over a million dollars to Trump and Republican campaigns. They got billions back. That’s the transaction. That’s the one-party state at the federal level, flowing directly into Louisiana.

2

Just read a piece from the team at Camber Creek portfolio company Proof on the evolving state of authentication…and it hit home. For years I’ve been thinking “passwords are dead,” but also lamenting the difficulty navigating the labyrinthine world of advanced security. This blog lays out exactly what’s replacing passwords in a way that’s practical, not just theoretical. Five methods stood out: - App-based authenticators finally killing SMS (good riddance, SIM-swap nightmares) - Continuous authentication that watches behavior during the session (the “someone stole the cookie but can’t act like you” protection we’ve all been missing) - Adaptive authentication that doesn’t treat grandma on her usual iPad the same as a login from a Tor exit node in a new country - MFA done right (and then taken further with verified digital certificates) - Biometrics with actual liveness + document tying, so it’s not just “a face” but your verified face What really resonated was the reminder that the strongest systems aren’t about piling on more factors; they’re about using the right factor at the right moment. Risk-based, context-aware, and (crucially) user-respecting. I’m biased but Proof’s angle of anchoring everything to high-assurance digital certificates feels like the missing glue. It’s one thing to authenticate someone at login; it’s another to carry that verified identity forward through the entire session, signature, or transaction. This one is worth the read. Link in comments. #venturecapital #cybersecurity #fraud

23

6 Comments

The future of GTM is autonomous—and Patrick Thompson’s recent article makes that crystal clear. What’s always impressed me about Patrick (even before Clarify was founded) is his systems thinking. For years, we’ve talked about the risks and potential of orchestration, testing, and observability—not just as DevOps principles, but as foundational layers for revenue. His piece on The Autonomous GTM Stack lays out the next evolution: AI agents executing across CRM, web, ads, and enrichment platforms—at speed and scale. But here’s what we both know: You can’t scale autonomy without observability. We've been on the observability train for a few years, and are building the test and observability layer for that future. If you haven’t read his article with Madrona yet, start there. Then ask yourself: Can your stack see what it’s doing? #AutonomousGTM #GTMLeadership #RevOps #RevenueObservability #Clarify #StackMoxie

4

An LLM will happily transfer money to the wrong account based on a wild guess. These are the top two issues I find with LLM "tool calling", day in, day out: 1. False-positive activations. - Customer says, "I've been thinking about the upgraded plan, it looks interesting," and 1-2 times out of 10, the model decides to call upgrade_plan(). - "This is outrageous! I'll cancel my subscription over this!" and immediately calling cancel_subscription() without "understanding the game." 2. Parameter guessing, which is particularly problematic with critical parameters that should only come from the customer. So the simple statement "I need to send money to a friend of mine" all of a sudden leads to transfer(recipient="john_smith", amount=1000) just because a similar transaction record is contained in the retrieved context for the model call. And yet, most applications written today still treat agentic tool calling as: "Here are your tools, you know when to use them... go nuts, baby!" Then people are shocked when the agent actually goes nuts. So here's how you can fix it.  👉 Tools shouldn't be available for evaluation all the time. They should only even be considered when specific conditions are met in the conversation. This is exactly why, in Parlant, tools are bound to guidelines. So that, first of all, that transfer() tool is only even considered when the guideline condition "the customer explicitly asks to transfer money" is met. Before that, the tool doesn't even exist as far as the agent is concerned. That's the first part of the solution. The second part is that, for critical parameters, you can annotate them with useful constraints, such as [source="customer"], forbidding the agent from trying to infer them. So if the customer hasn't explicitly provided that value, the tool won't run, and the agent will ask for it or at least confirm its guess before running it. At the end of the day, here's the difference in your agent's behavior: In the traditional "go nuts" approach: "I think the customer wants to transfer $1,000 to John Smith. Let me go ahead and do that!" But when you use Parlant, it's: "I can see that the customer wants to transfer money, and I was instructed to consider this tool for this use case. While they specified the recipient, they still didn't say the amount, so I need to ask for it." And this is what production-ready tool calling looks like. Not guessing or going nuts. Yes, it's more effort to set up than just letting the LLM guess. "But wait," you say, "No, no, no. The LLM should be able to guess it correctly. If it doesn't do it today, we'll keep building in the same way because GPT Z will get it right, so we'll just wait for it to be released. Then we'll release our agent." Heard this *too many* times 😃 I'll leave it as an exercise to consider, "Will any model ever be able to guess it all correctly and consistently? Will we ever not need this encoded human guidance in the loop?" (hint: first principles)

32

3 Comments

In terms of LLM use cases, one of the highest value add uses is software coding, and in this application, the proprietary models dominate: "proprietary models still handle the bulk of coding assistance overall (the gray region), reflecting strong offerings like Anthropic’s Claude." This is especially important because coding is now the dominant, token-weighted use case — in early 2025, coding represented 11% of token volume, but toward the end of 2025, this share exceeded 50%. The following exhibit from the report highlights proprietary model dominance in coding use cases:

1

We are moving quick with our Railroad Commission of Texas MCP. We should have all datasets by the end of the week, sans unstructured data. Here’s what’s avalible right now. searchWells searchWellsByArea searchInspections listPipelines searchPipelineSegments searchGasProcessingPlants

6