Raj

14+ minutes on hold with United Airlines. 90 seconds of my time. Week 2 of building AskRanger.com in public. After a March family trip to Hawaii, I realized I forgot to request a refund for the non-functional Wi-Fi on our United flight. A month passed. Trying online failed, which meant calling, which meant hold music, which meant it would sit on my to-do list for another month. Instead, I texted Sara (my personal agent from AskRanger): "Call United and get a refund for wifi that didn't work on flight UA694." Sara called United. Navigated the automated phone menu. Waited on hold for 14+ minutes. Got a rep on the line and explained the situation. The rep needed my MileagePlus number to verify my identity. Sara texted me: "They need to verify your identity - want me to patch you in? Text JOIN or tell me what to say." I replied JOIN. Sara connected me directly to the rep and dropped off the line. I confirmed my info, and the refund was approved. A full credit of 800 miles back the next day. My total time: 90 seconds. Sara handled the 14+ minutes of hold music I was happy to miss. Thank you to our early users and everyone on the waitlist - we're opening more spots as we harden the product. Still in private beta. Text GO to (510) RANGER1 or DM me for an invite code.

This is the autopilot thesis written cleaner than I've seen it elsewhere. The wedge is consumer - real-time outcomes that prove the AI works. The destination is vertical services where the seat-to-result transition is most painful (medical AR, real estate ops, industrial sourcing, anywhere humans currently sit on hold for hours). Same infrastructure, different buyers. The interesting question is which vertical pulls the consumer infrastructure into managed services first.

5/ Stop selling the seat. Start selling the result. Full piece on when this transition makes sense, what changes, and why waiting is the bigger risk: emcap.com/thoughts/shoul…

1/ Two weeks ago I took a walk with the CEO of a SaaS company with hundreds of millions in ARR. He told me he's pivoting the entire business to a services model. Can't share who yet. Soon.

Deployment is what matters. Deployment is not solved.

Hear what we @DynaRobotics learnt from the bitter lesson in the past 10 months in production

@jasonlk You’re right at SaaStr’s or later stage. As a solo founder, build is still my preference… preserving runway matters, and agents let one person do what used to take a team. The buy/build line moves with company stage and team size, and it will towards buy more as products mature.

Maybe but I don’t think so We don’t have the time to make those workflows as strong and reliable and bulletproof. And there is only so much time. We just don’t have any more time to build any more core, complex agents ourselves. Almost always cheaper to buy than build, if you include soft costs. IF it’s actually available to buy.

I just pulled SaaStr’s Replit bill for last month. Our two AI VPs cost us $254 for the entire month. Combined. - 10K (our AI VP of Marketing): $94.51 - Qbee (our AI VP of Customer Success): $159.55 Two human VPs at the same level would run $500K-$800K a year, all-in. A few things before anyone takes shots at it: The $254 is real. But it’s only true for agents we built ourselves on Replit. Our third-party AI agents (Artisan, Qualified, Agentforce, Momentum) run $25K+/yr each. Apples to oranges. The $254 also doesn’t include the soft costs. We update 10K daily. Qbee a few times a week. Budget as much as 0.2 FTE of human attention per production agent. The “deploy and forget” pitch is wrong. Buy don’t Build / Vibe it yourself … if you can. If it exists and meets your needs. But even with daily improvements and maintenance, managing these two AI VPs is still less time than managing one human VP, much less two. No 1:1s. No comp conversations. No PIPs. No backfill recruiting. The maintenance IS the management. Every hour tuning the app directly ships output. A few more numbers from the bill: – Total Replit bill: $2,324/mo – Runs 6 production agents + 14 published apps – 1.9M+ requests served/mo – Whole stack costs less than one mid-tier SaaS tool We run SaaStr with 3 humans + 20+ agents. Revenue went from -19% YoY to +47% YoY in the same period. The full cost of running an agent-augmented business in 2026 is more like $50K-$200K/month, not $254. Still dramatically less than the human-only equivalent. But the headline is real: two AI VPs. $254. One month. That’s the new baseline. (And we’ll show you how to build your own AI VP Marketing + AI VP Customer Success at SaaStr AI Annual 2026, May 12-14 in SF Bay. Join us!!)

@levie Agree it’s a utility. The interesting design question is whether utilities benefit from relationship-style interfaces (memory, continuity, name) even when users intellectually know they’re tools.

In general, we should treat AI like a utility, not like a being. The more we confuse what AI is the more we will make ourselves go crazy with analogies that will never fully hold true.

I think Founder-CEOs do care, but most are 5 layers deep in their org and the data has to climb through everyone before it reaches them. At my last company (Flock, sold to Paychex), I'd ask eng for a dashboard and wait 2 weeks. Now I spin one up in hours with an agent. The CEOs who can navigate their own org and build their own business monitoring will eat the ones who wait.

@hrhraj I think the reality is no one cares Customer Success is dead, and some churn is inherent to all models They should care. So much cheaper to keep a customer than acquire a new one. But most don’t

We stealth-churned off Notion months ago. We didn't realize it. We didn't tell Notion. It's just ... our agents have no need for it. They're doing the work now. Stealth churn is the under-discussed agent risk. Usage drops to zero quietly. The renewal email never even goes out.

@andrew__reed From Barbarians at the Gate to kk-arrrrrrr..... the evolution is complete

what's the best pirate equity firm? kk-arrrrrrr

@naval All movement is not progress....

“What did you build this week?” is the new “what did you get done this week?”

@andrew__reed Or like this ... Senedd, Wales

Saved me personally about 15 mins of hold time with United for a wifi refund. The refund value was $8 / 800 miles... but my 15 mins - priceless. Also sourced the viral 'mystery dumpling' toy by calling 4 stores.. found when it's in stock.. saved me 32 mins, and now my daughter thinks I am a hero.

@hrhraj How much money do you save everyone?

Describe your product in exactly one sentence. No buzzwords, no fluff, just the core value. If I can’t understand your business in ten seconds, I’m not investing. Hit me & i'll be in your DMs

The dynamic interface piece is interesting. We've been thinking about this... instead of returning a list of plumber quotes as text, generate a comparison page with their schedules, ratings, and approval buttons. The agent already has the data; the interface just needs to be situational. But hard to do over sms/text....

Some observations on agents from a lot of time spent installing, using, debugging and building things with OpenClaw, Hermes, etc. - Most people who rave about OpenClaw, HermesAgent etc. have not actually used it. Facts. - Agents are developer products today. They are so so so far from being consumer grade in terms of delight, simplicity, reliability etc. - I wish the agents generated more dynamic interfaces. Don't give me chat. Give me interactive, dynamic apps/webpages. I suspect this will be a next big vector of innovation. - The agents are ultimately a wrapper to code-generation and tool-calling....not to "generative AI". This is important to grok. These agents are doing most of their "work" through iterative tool calling. - Once you understand the core loop behind agentic AI, old school "chat AI" doesn't make sense. The idea of an "always-on" agent that actually adapts, learns and maintains state is a LOT more compelling than a one-off chat prompt. The latter feels much more like a search query. - So much of mangling with agents today is basically setting it up to use browsers and/or your computer. And there are so many constraints! - Installing and running openclaw feels a lot like running Linux in the early 2000s. You spend a lot of time compiling device drivers, setting up configurations etc. You feel a real sense of accomplishment from jerry-rigging everything together but it is a LOT of work to get to the point of utility

I tried to for a grape harvest project I was exploring… the gripper kept squashing the grapes.. had some left over Filaflex from my amazing hand project so attempted to build this for the gripper, but pivoted to using two stereo cams so it could grab by the stems so force didn’t matter.

@hrhraj Wow cool you actually made this?

"eFlesh" (e-flesh.com) sounds like porn site, but its a robot tactile sensor developed at NYU 😊 You'll need a magnet, a magnetometer sensor board ($5 on adafruit for 1 x 1), and 3d print the skin. This is a real low-cost and versatile way to make grid pressure sensing.

@philfung This worked well for me for tactile amazon.com/dp/B074QLDCXQ and for capacitive (WWZMDiB 6Pcs TTP223 Digital Touch Sensor)

Q: what is the best low-cost tactile sensor to add to a robot gripper?

@pham_blnh Very cool. I did something similar with a NEMA 11 w/TMC 2209... works well and much quieter for my use-case but your approach has less failure points and simpler.

i bet you haven't seen a SO101 mounted on a wall like this before if you want to do the same, here is LeSlider: github.com/pham-tuan-binh… i built it cause i wanted something that can cover my whole desk for tasks like organizing and cleaning i originally wanted to have a belt system like what 3D printers have, but i was too lazy and used a pinion/track with another sts3215 so: > the extra motor shares the same bus as the rest of SO101 > you can have arbitrary length of track > really cheap and easy to assemble and control it turned out better than expected with this, i'm gonna train a model to pick random stuff up across my table and put it into a bin at the end of table (realistically using yolo to scan table, two policies, one for picking up objects, one for dropping)

@MattHartman Most flexible… OpenSCAD with python (use any frontier llm and plug in code). But a bit of a learning curve. A little setup and this is also good GitHub.com/eqrthtojake/te… @ForgeCAD is easiest… no setup.

I want to vibe-CAD my next robot. What is the best vibeCAD software to start with that will ultimately help me pick motors, 3d print etc. What's out there that people have tried & like?