Andrew

@WorkWeave | Prev: @causalhq, @MIT | he/him

[Andrew @AndrewChurchiII]

@telnyx @davidcasem Probably the most interesting part to me: despite AI doing all this work, Telynx has 40 open engineering roles right now - they are hiring more engineers, not fewer! (9/9)

Full write-up: https://t.co/4UfP23v2k3.

[Andrew @AndrewChurchiII]

This week I sat down with the CEO of @telnyx, @davidcasem, and it might be the most interesting AI-native engineering org I have talked to all year. (150 engineers, 1400 bots, 0 engineering managers(!))

They rebuilt how work enters the org, who does it, and how it ships. (1/9)

[Andrew @AndrewChurchiII]

@telnyx @davidcasem 𝗕𝗲𝘀𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝘀 𝗱𝗼 𝗻𝗼𝘁 𝗴𝗲𝘁 𝗵𝗲𝗮𝗱𝗰𝗼𝘂𝗻𝘁. 𝗧𝗵𝗲𝘆 𝗴𝗲𝘁 𝗰𝗼𝗺𝗽𝘂𝘁𝗲. Seniority is measured by ROI per unit of AI resources, not by the size of your team. (Weave helps with this!) (8/9)

[Andrew @AndrewChurchiII]

Citadel doesn't give its best traders bigger teams, it gives them more capital. I think engineering is heading somewhere similar.

Seems like a random connection, but let me explain!

First, some key context: multi-manager hedge funds like Citadel and Millennium are made up of groups called pods. Each pod is a small, semi-autonomous unit with its own P&L. The firm provides the shared infrastructure, and capital flows toward the pods generating the best risk-adjusted returns.

Managing one of these pods is *extremely* lucrative - the contracts rival professional sports players for annual salary, and the competition between different firms is intense!

Now swap capital for compute.

In an AI-native engineering org, your most valuable resource isn't headcount anymore, it's tokens and the systems around them. So what happens if you organize engineers into small pods, give each one a budget of AI resources, and measure the return they generate on it?

The reward structure starts to look like a fund too. Your best people get more resources to direct, not more reports to manage. Seniority becomes ROI per unit of compute rather than team size.

There are real problems. Engineering output is genuinely hard to measure and credit, much harder than a trading P&L.

And unlike a fund where managers can place overlapping bets freely, engineering teams have to coordinate to avoid building the same thing twice. The pod model assumes independence that engineering doesn't fully have.

I don't think this is the full picture. But the direction feels right, and I think there will be some convergent evolution between the two fields.

This is also why we are building Weave to solve the engineering measurement problem. So you can allocate your resources in the most effective manner.

[Andrew @AndrewChurchiII]

We have 5 full-time engineers at Weave and we regularly run 20+ concurrent work streams. This has a massive effect on our throughput, and queuing theory explains why. Stick with me on this one!

In queuing theory, wait times don't scale linearly with how busy a system is. If you take the simplest possible model of a queue (M/M/1 for the nerds) you see that as utilization approaches capacity, wait times go to infinity on a hyperbolic curve, not a straight line. At 80% utilization, the average wait time is 4x slower than at 50%. At 90% it's 9x worse. The math is (wait time = utilization / (service rate - arrival rate)), but systems that are almost full behave way worse than systems that are only somewhat full.

The flip side is that adding parallel capacity produces disproportionately large improvements. Going from 1 elevator to 2 in a building doesn't halve average wait times. Depending on traffic patterns it can cut them by 10x or more, because you've moved the system off the steep part of the curve.

Engineering teams have historically been single-threaded. One engineer, one task, and however many engineers you have is your ceiling for concurrent work. Coding agents changed that for us. Our engineers now run 2-5 workstreams simultaneously by having agents handle execution while they focus on architecture and review.

So when our 5 engineers go from running 5 concurrent threads to 20+, queuing theory predicts the throughput gains should be nonlinear. In my experience that's been true, and our own metrics agree. Work moves through our system faster than the headcount alone would suggest is possible!

[steven @syriansigma]

https://t.co/fnwm34H4Ia

[Andrew @AndrewChurchiII]

The teams that win in 2026 won't be the ones who code the fastest. They'll be the ones who make better decisions about what to build and who to build it with.

[Andrew @AndrewChurchiII]

We track 10,000+ engineers on Weave. Junior engineers with product sense are starting to outperform seniors with decades of experience. That's just one of many trends I think will dominate in 2026:

[Andrew @AndrewChurchiII]

𝟰. 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲𝘀 𝘀𝗲𝗲 𝗿𝗲𝗮𝗹 𝗿𝗲𝘀𝘂𝗹𝘁𝘀
The companies that adopted AI aggressively in 2025 will start seeing actual business outcomes in 2026. Not just "engineers are more productive" but real improvements to revenue and customer satisfaction.

[Andrew @AndrewChurchiII]

If you're not curious about the code your AI is writing, you're not actually engineering. You're just being a middleman.

[Andrew @AndrewChurchiII]

After reflecting on the 250 interviews I have done in the last 6 months I've noticed I subconsciously filter out candidates who lack one trait:

Curiosity.

[Andrew @AndrewChurchiII]

Being curious is uncomfortable. It means constantly asking "why does this work this way?" instead of just accepting that it works. But that discomfort is what separates good engineers from the rest.