Ten Engineering Leaders All Agree That The Real Bottleneck Sits Outside the Code: Here's Why

Ten TechPod Talks Guests. One Fix You Can Run This Quarter.

Every engineering team can now generate ten or a hundred times more code, and almost none are shipping ten times the value. That gap forces a decision that has nothing to do with which model or coding assistant a team adopts. Instead, it asks engineering and business leaders where the real constraint moved once code generation stopped being one.

Ten engineering and technology leaders answered a version of that question across the first ten episodes of TechPod Talks. They came from industries across the board, spanning gaming, media and entertainment, consumer technology, enterprise software, cybersecurity, and organizational consulting. These are the individuals who run their own engineering orgs and security teams, sit on the leadership side of platform decisions, and even help advise the executives who make these calls. However, several of them disagreed with each other directly, on air, without knowing what the others had said.

On our last episode, Stephen Koza, CEO of EverOps, returns to the premise he raised back in Episode 2, arguing that coding stopped being the bottleneck and everything else became the constraint. He plays the ten answers back-to-back and names the pattern the guests reached without coordinating, revealing that the bottleneck kept landing more on organizational terrain and almost never on technical ground.

Read on to see where each of those ten leaders located the constraint, where they contradicted each other, and which of their fixes a VP of Engineering or CTO can put to work this quarter.

The Bottleneck Relocates to Whatever a Team Fixed Last

Robert Gonzalez, VP of engineering at SugarAI, described what happens once a team removes its coding constraint. Accelerating one part of an AI-driven workflow simply moves the bottleneck to the next part that has not caught up, whether that is deployment, go-to-market, or the business functions waiting on engineering's new output. 

This means that an engineering leader who accelerates coding without ensuring deployment and go-to-market are equally covered by AI ends up building a fast lane the rest of the business cannot use.

"Accelerating one piece of a puzzle just moves what happens in a bottleneck situation to a different part of that puzzle."

Francisco Trindade, VP of product engineering at Braze, puts a number on what that relocation looks like at scale. Picture 100 engineers each running five agents. That comes to 500 agents working against a codebase that has grown to 2 million lines, and holding that codebase coherent as it grows becomes the actual engineering job. 

Chris Harden, director of production verification and AI liaison at Unity, locates the same relocation on mature products, where agents introduce slop that a team inherits as years of tech debt if nobody verifies what shipped. His fix reframes the engineer's job as orchestrating the agents and verifying what they produce.

Agile Metrics Built for Scarce Code Do Not Survive Cheap Code

Robert Gonzalez simultaneously supplied the season's clearest example of a metric that stopped measuring anything useful. He explains that story points were built to estimate time and complexity when both were expensive, and as agents absorb more of the coding work, that estimate is collapsing toward a factor of one, with throughput becoming the number worth tracking instead. To help combat this, he built a Scrum Master agent that plans sprints against a team's true capacity, tracks completion and rework, and produces the daily and sprint-level reporting that used to take his staff days to assemble across his 12 to 13 Scrum teams.

The same system let him run an AI-assisted audit of 13 months of historical tickets, checking whether his teams estimated consistently. The variance he found argues for standardizing the measure before retiring it. One team averaged 1.1 story points per ticket, and a comparable team averaged 3.2 story points per ticket, for work that looked the same on paper. 

The critical takeaway here is that a leader deciding whether to retire story points this quarter needs that kind of audit first because a factor-of-one throughput number only means something once every team counts a ticket the same way.

Cost Ownership Decides Whether Savings Survive Growth

Tom Kershaw, SVP and CTO at BMI, watched request volume jump 25 times in a single month during his time at Rubicon Project. He also pointed to Deezer's own public numbers as the clearest illustration of volume outrunning value, noting that with roughly 60% of uploaded songs AI-generated against roughly 0.03% of plays, it created a flood of output that almost nobody consumes.

Mark Sass, Senior Director, Platform Engineering at SolarWinds, picked up the same thread with the budget version of the problem by claiming that cost optimization is consistently the first work that gets dropped when a team is focused on shipping to customers. This is because research takes time, even when a tool hands you an easy-looking win like a cloud savings recommendation.

Patrick McKinney, Chief Information Security Officer at Turing, ran a version of the same math against his own security operations. Standing up an agent-driven SOC came in at roughly one-fifth of what staffing an in-house team around the clock would have cost. That gap turns the conversation about bringing people back into a concrete budget line a CFO can approve this quarter. He kept a human detection and response team monitoring the platform alongside the agents, because the savings only hold if a person still owns the judgment calls the agents cannot make.

"AI should be supercharging humans, not necessarily just replacing jobs one to one."

That single line describes how Patrick built his own team. He hired lean and organically as functions such as infosec, cloud, and fraud detection needed dedicated coverage, and he assigned each new hire the job of defining which processes AI could run and which still needed a human sign-off. 

Today, a leader running the same math on a high-volume manual process should expect the same split, including a large cost advantage from automation and a smaller, more senior human team that owns the exceptions.

AI Pilots Stall on the Org Chart Before They Stall on the Model

Dr. Janet Sherlock, founder and CEO of Org.Works, attributes most AI pilots that never scale to organizational structure. The industry's current answer is what she calls the "great flattening," broad middle-management cuts at firms like Meta, Amazon, Microsoft, Intel, and Expedia aimed at shortening decision paths. She recommends thinning vertically before flattening horizontally, removing coordination layers that sit between teams that should have been talking from the start. Her model for what replaces those layers has three parts: 

1. A center of enablement owns platform and data infrastructure
2. Federated teams hold the domain expertise and ship the use cases
3. Democratized data and AI tools stay available across the organization

Chris Robertson, VP of cloud operations and IT at Arlo, puts a number on the same organizational drag from the reliability side. Getting marketing, product, and the executive team to agree on a goal eats 40 to 50% of any reliability program, before anyone touches the architecture. He treats that agreement as a prerequisite a leader has to secure before a technical fix can even be evaluated for success, since a team can hit every engineering milestone on a reliability push and still watch the program stall the moment product wants to ship a feature the fix was supposed to delay.

Sherlock's and Robertson's numbers describe the same failure from two different vantage points. One measures the coordination layers a pilot has to cross before it reaches production, and the other measures how much of a technical program's total effort goes to securing agreement before any building starts. 

"If you can't ask a good question of the tools, the tools will tell you exactly what you want to hear and they will drive you straight off a cliff."

So, a leader who treats either number as a rounding error is the leader whose AI pilot or reliability push stalls next.

The Guests Who Disagreed Backed It With Specifics

Patrick McKinney draws the season's hardest line on vibe coding, stating:

"AI is not writing quality code. And if anybody says that it does, then they're 100% wrong."

Meir Wasserman, Head of Engineering at 2K Technology, reached a similar conclusion independently. On his episode, he tells Stephen that more generated code isn't automatically better, and that he would rather see his team generate less code with a smaller attack surface. He also raises the junior engineer pipeline question directly, admitting he has mostly worked through his own concern that AI would remove the entry-level reps junior engineers need.

Francisco Trindade backs him with his own history as a junior engineer, when his team spent a week of six people standing up a build server before anyone could commit code, a cost and a learning curve that got automated away without erasing the junior track. Jose Gonzalez, VP of product at Ticketmaster, argues for hiring a person who can redefine a role once they are in it. He treats a new hire's job as reshaping the space they land in, since the job posting was written before AI changed what the role needs to do.

None of these four guests coordinated an answer but still shared similar perspectives. Patrick and Meir landed on the same skepticism about vibe coding from a security lens and a game-studio lens. Francisco and Meir landed on the same optimism about the junior pipeline from a 2026 engineering-leader vantage point and a memory of building infrastructure by hand decades earlier. 

Jose's hiring philosophy answers a question none of the other three were asked. That range, arriving at overlapping conclusions without a shared script, is what gives the disagreement its weight for a leader deciding how much to trust either position.

What These Ten Episodes Add Up To

Google's own 2025 DORA report surveyed nearly 5,000 technology professionals this year and reached a parallel conclusion, stating that AI amplifies whatever a team already has, strong practices and weak ones alike.

Season 1 of TechPod Talks reached parallel conclusions all of the conversations, without ever comparing notes. These are the ones you should pay attention to: 

• The bottleneck never disappeared. It just shifted from code generation to review, deployment, go-to-market, and organizational agreement, and it will keep moving to whatever a team traditionally fixes last.
• Metrics built to price scarce code, especially story points, stop measuring anything once agents make code cheap, and a team needs a shared baseline before it can trust a throughput number.
• Cost optimization is the first work volume growth crowds out, and it only survives when the teams generating the spend can see and own their own number.
• Token costs passing human costs turns hiring people back into a real budget conversation, provided the humans hired keep the judgment calls agents cannot make.
• Org structure decides whether an AI pilot scales far more often than the model or the tooling does.
• Getting marketing, product, and the executive team to agree on a goal is 40 to 50% of any hard technical initiative, before evaluating any architecture.
• The guests who pushed back on the season's premise did it with specific numbers and specific experience, on vibe coding, on the junior pipeline, and on what a new hire should be trusted to redefine.

How You Can Apply These Frameworks This Quarter

1. Map your bottleneck one hop downstream: Compare code review, deployment, and go-to-market throughput against your current code generation throughput, and fund the widest gap.
2. Audit your estimates before you retire them: Run an AI-assisted review of the last 12 to 13 months of tickets the way Robert Gonzalez did, and standardize what a story point means across teams before you switch to a throughput measure.
3. Push budget ownership down to the teams generating the spend: Give each team visibility into its own number and make its manager the first line of accountability, so a cut survives longer than six months.
4. Model your highest-volume manual process against an agent-driven build: Run the cost comparison Patrick McKinney ran on his SOC before shifting headcount, and staff the resulting team to own the judgment calls agents cannot make.
5. Diagnose a stalled AI pilot as an org chart problem first: Use Janet Sherlock's center of enablement, federated teams, and democratized data model to find the coordination layer slowing adoption before you re-evaluate the tooling.

Ten Episodes Down and a Playbook That’s Ready to Run

These ten conversations emphasize that the constraint on shipping value with AI is almost never the model or the code it writes. That throughline becomes work the moment a VP of Engineering or CTO decides which stage to fund next, how to price a ticket once story points stop meaning anything, who owns a cloud budget, and how many coordination layers sit between a pilot and the teams it is supposed to serve. 

EverOps built the AI Opportunity Assessment and the AI Quick Start to answer exactly those questions inside a working roadmap, backed by the same platform engineering practice several of this season's guests described building internally. 

If Season 1’s ten conversations point toward a decision your team is facing now, reach out to our team. We will help you run it down.

Keep up with TechPod Talks

Subscribe on Apple Podcasts, Spotify, YouTube, or the EverOps podcast page. 

Follow EverOps on LinkedIn for upcoming Season 2 guest announcements.

Go back and listen to every episode we’ve referenced: 

• Mark Sass, Senior Director, Platform Engineering at SolarWinds
• Francisco Trindade, VP of Product Engineering at Braze 
• Dr. Janet Sherlock, Founder and CEO of Org.Works 
• Chris Robertson, VP of Cloud Operations and IT at Arlo 
• Meir Wasserman, Head of Engineering at 2K Technology 
• Robert Gonzalez, VP of Engineering at SugarAI 
• Chris Harden, Director of Production Verification and AI Liaison at Unity
• Jose Gonzalez, VP of Product at Ticketmaster 
• Tom Kershaw, SVP and CTO at BMI 
• Patrick McKinney, Chief Information Security Officer at Turing

Frequently Asked Questions

What is TechPod Talks?

TechPod Talks is a podcast from EverOps featuring candid conversations with the engineering and technology leaders behind major platforms. Stephen Koza, CEO of EverOps, hosts each episode and presses guests on how they are running their organizations through the AI transition.

Who is this podcast for?

The show is for VPs of Engineering, CTOs, and directors and senior managers whose findings travel up to the leadership team. Episodes focus on the decisions leaders make around architecture, cost, staffing, and organizational structure.

Where can I listen to TechPod Talks?

Episodes are available on Apple Podcasts, Spotify, and YouTube, along with full transcripts on the EverOps podcast page.

What did Season 1 conclude about the engineering bottleneck?

Across ten episodes, guests independently identified the real constraint on shipping AI-generated code as something other than the code itself. Episode 13 plays those ten answers back-to-back and names that shared pattern the season's throughline.

Who were the guests in Season One?

Season One featured Mark Sass, Francisco Trindade, Dr. Janet Sherlock, Chris Robertson, and Meir Wasserman across its first five episodes. Robert Gonzalez, Chris Harden, Jose Gonzalez, Tom Kershaw, and Patrick McKinney closed out the back half. Together they led engineering, security, and product functions at companies including SolarWinds, Braze, Arlo, 2K Technology, SugarAI, Unity, Ticketmaster, BMI, and Invisible Technologies.

How much does an AI-driven security operation actually cost compared to staffing an in-house team?

Patrick McKinney of Invisible Technologies found that standing up an agentic SOC platform cost roughly one-fifth of staffing a round-the-clock in-house team, while keeping a smaller human team in place to monitor the platform and make the judgment calls agents can't.

How does this episode connect to EverOps' work?

EverOps helps engineering leaders run the same decisions this season's guests describe, through the AI Opportunity Assessment, the AI Quick Start, and ongoing platform engineering work that keeps a growing codebase coherent as more of it gets written by agents.