Scott Dietzen thinks the job of a software engineer is about to change shape, and he has a name for the new one.
“the role of the developer is going to move to being a tech lead for a team of agents.”
Scott Dietzen, at 2:37
Scott, “Dietz” to almost everyone, was the chief executive of Augment Code, the company building AI for software teams with large code bases, from April 2023 to July 2025, and he has stayed on its board (LinkedIn). Before Augment he ran Pure Storage from September 2010 to August 2017, as it grew from 15 employees to nearly 2,000 and from zero to a billion dollars in revenue through its IPO, then chaired its board, and he is its vice chairman today (Augment Code). Before that he was president and chief technology officer of Zimbra, which Yahoo bought in 2007, and he spent two years at Yahoo running applications. His PhD is in computer science from Carnegie Mellon, where Geoffrey Hinton taught his introductory AI class. We recorded in 2025, after Augment’s ISO 42001 certification at the end of May, and published the episode on July 29, 2025, with 47 words of show notes. This page is the rest of it: a section for each part of the conversation with the sources beside it, nine figures, seven clips from the tape, four videos, two posts from X, the full transcript, and four tools: a quiz, a scorecard for the AI coding assistant on your team, a calculator for how long a task your agents can take on next year, and a timeline.
A disclosure before we start. YSecurity ran Augment’s SOC 2 Type 2 and ISO 42001 programs, I wrote Augment’s post announcing the ISO 42001 certification (Augment Code), and Scott thanks us at the end of the tape. So his claims are labeled as his below, and the public record sits beside them. My co-host Sasha Sinkevich is on this one too.
The anti-vibe coding company
I asked Scott for the short version of Augment and its mission, and he gave it to me in one line.
“You can think of us as the anti-vibe coding company, if you like.”
Scott Dietzen, at 0:27
Vibe coding is the name for describing what you want to an AI and accepting the code it writes without reading it closely, which works beautifully for a weekend project. Augment was built for the other end of the job. In his telling, the team first looked at large language models about four years before we recorded and saw their potential for the hardest kind of software work, the large enterprise systems with “many tens of millions of line code bases” that are, in his words, “the ones that run the world.” Then came the line that is the whole thesis.
“In fact, you could say there’s no simple change in a hundred-million-line code base, right? It just doesn’t exist.”
Scott Dietzen, at 0:27
He has written the same idea down. In a January 2025 post on Augment’s blog he put the size of the problem in numbers, from more than 30 million lines in Chromium to more than 90 million at Uber and more than two billion in Google’s largest repositories, and gave it an image: if programming is a toy airplane, software engineering is “constructing an Airbus A380 and continually remodeling it in flight” (Augment Code). What makes an AI useful in that setting, he said on the tape, is exactly what makes the setting hard for people: “AIs are actually phenomenally good at managing detail across really large pieces of information,” so they make “a great co-conspirator” for the engineers who have to evolve those systems.
The reason to bother is that the software in those systems isn’t as good as it should be, and he didn’t soften it.
“because most software in the world disappoints, right? It’s not as reliable as it should be. It has security vulnerabilities. It doesn’t have all the features that we want. It can be hard to use.”
Scott Dietzen, at 0:27
He put a price on it from memory: “We’ve had, I believe last year, $2.5 trillion lost in the US economic output solely due to software failures.” The record makes his point with a slightly smaller and slightly older number. The estimate he means is the Consortium for Information and Software Quality’s 2022 report, which put the cost of poor software quality in the United States at “at least $2.41 trillion” that year, with technical debt alone at about $1.52 trillion (CISQ). His own January 2025 post cites the same report as $2.4 trillion. Either way, it is about a trillion dollars more than the world spends on software engineering salaries in a year, which Augment’s launch release put at more than $1 trillion, citing Gartner (Augment Code).
I tried to say it back to him in one sentence, “The anti-vibe coding co-conspirator for all engineers everywhere working on, like, ugly, large, legacy code bases connected to big enterprise systems.” He took it. “That’s it. In a nutshell.”
A tech lead for a team of agents
Then I asked where Augment was on its journey, and his answer was the year’s big change.
“So, the big disruption that’s happened over the course of this year is agents.”
Scott Dietzen, at 2:37
The best developers, he said, are already ready for it: “the most sophisticated developers are already very comfortable in that kind of role, because they do it today for humans.” A tech lead hands out tasks, checks the work and keeps the design in their head. With agents, the same person can have “a suite of agents at your beck and call” for the less fun, more tedious work, “and being able to tee up a bunch of them in parallel and then check on their work.” He wrote the same promise into Augment’s agent launch post that May, that agents would “turn every engineer into a tech lead, commanding a team of autonomous AIs,” and named the catch in it too: when agents misunderstand what you meant, fail the task or break existing systems, “they become a liability rather than an asset” (Augment Code).
That is also where the company’s name comes from.
“[We] named the company Augment because we think we are still a long ways from not needing human insight and intellectual guidance to yield ideal results.”
Scott Dietzen, at 2:37
The decisions that stay human, in his list, are the quality of the code, the architectural approach, the data model, the database, which cloud service, whether to go to microservices, and above all checking the results. Augment’s launch release in April 2024 had Dion Almaer, then its vice president of product and a guest on episode 43, say it the same way: “The best software comes from augmenting developers, not replacing them” (Augment Code).
He was bullish on the size of the change, by his own estimate: “we’re no longer talking about 30, 40% improvements, right? We’re talking about 300, 400% productivity improvements with the introduction of agents.” Those are his figures, from Augment’s customers as he saw them, not a study we could check. Sasha added the reading of the name that I’d put on a slide, “It’s the extension of the human force, hence the augmentation of the workforce,” and steered us toward the part nobody talks about, the data that goes into an agent: “It’s important to provide really good data as an input in order to have a good output.”
Every step up the abstraction hierarchy
Sasha’s point about data is the one Scott spends the most time on, because it is the one Augment was built on. He walked the history of AI coding tools as a ladder.
“These AI assistants started with code completions, where context was helpful, right? You wanted to reuse code rather than add new code.”
Scott Dietzen, at 4:27
Then chat, where you ask questions about a complex code base. Then next edit.
“Next edit is where the AI watches you make a change and then infers what else you might need to change along that journey.”
Scott Dietzen, at 4:27
Augment shipped Next Edit for VS Code in February 2025, for the change every engineer knows, where one new field means “updating SQL queries, tests, and type definitions to match” all over the code base (Augment Code). Agents are the next rung, and each rung asks for more understanding of the program than the one below.
“Agents need more autonomy, and that autonomy depends on their having knowledge of what’s happening inside the software in order to make the right call.”
Scott Dietzen, at 4:27
His claim is that the years Augment spent on that understanding paid off when agents arrived. He called it the context engine, built “to be able to handle large, complex code bases,” and said it was “now paying us back many times over in the land of agents, where our agents are vastly better than the alternative software engineering agents on the market.” The record has one benchmark from that spring. On March 31, 2025, Augment reported a 65.4 percent success rate on SWE-bench Verified, the top published score for an open-source agent at the time, and open-sourced the approach (Augment Code). Augment Agent itself launched on April 2, built for anything “from a brand new app to a monorepo with over 100k files” (Augment Code).
He finished with the kind of work it frees people from: spec out a new implementation, then “crank me out a bunch of unit tests and run them,” do “a quick security pass,” and when something breaks in the build pipeline, “you can have an agent troubleshoot it, so that humans don’t have to go through a lot of the manual, tedious work.”
Chat with the code base
Security was my way in. When we run a red team exercise or a penetration test, we always ask for the code.
“What we’ve seen is, like, during a red team exercise or a penetration exercise, we always prefer to give those white hat hackers access to the code base of whatever it is that they’re looking at.”
Jon McLachlan, at 6:23
Static analysis tools are part of that, and “Augment is certainly not a replacement for those things,” I said. But with an AI that understands the whole code base in the mix, a tester can ask for a very high-level, very detailed read of it and check for “paradigms that seem a little bit fishy or that might be able to be exploited.” Even before any agent does anything, the reconnaissance gets better: you can chat with the code base.
Scott took it somewhere broader. Most of the talk about coding AI is about productivity, he said, and “I don’t think we focus nearly enough on higher quality, more feature-rich software that results, right?” Then he named what is new for security.
“These large language models, they’re not deterministic. So, you know, we can’t necessarily test all the code paths and know where the vulnerabilities are. And we can’t try every possible prompt and know how the LLM is going to behave.”
Scott Dietzen, at 7:20
So, in his view, the defense has to use the same tools, to “further harden our applications against new security vulnerabilities that are going to come as just an inherent result of AI being an integral part of the software system.” He had a line for the security leaders listening, “CISOs everywhere are going to be challenged to make sure that these new systems are secured,” and a reminder that “attackers may bring AI to bear in trying to hack into our systems. So we’ve got to do, you know, a better job of defense than we ever have.”
This is the side of the conversation I’d underline for a startup. The same understanding that lets an agent change a hundred-million-line code base lets a tester read it, which means your code review, your penetration test and your threat model can all go faster and deeper. Our AI penetration testing guide has the practical version, and episode 84 with Jack Cable is about what gets missed when nobody reviews the code.
Proof points, attestation points
I congratulated him on the certifications, and I should say what I knew.
“As I understand it, Augment was the first coding platform to achieve SOC 2 Type 2 out there, and that was last year. And then this year, the very first coding platform to achieve ISO 42001 certification.”
Jon McLachlan, at 8:34
The record agrees on the dates and uses the company’s own words for the firsts. Augment received its SOC 2 Type II attestation on July 10, 2024, after an audit window from April 1 to July 1 “during which we did not have a single issue,” as Scott wrote that September (Augment Code), and Augment’s ISO post calls it the first dev tool to reach that standard (Augment Code). ISO/IEC 42001, the standard for an AI management system, followed on May 29, 2025, certified by Coalfire, whose release was headlined as the first for an AI coding assistant (Coalfire). We ran both programs, and the case studies are public: SOC 2 Type 2 in five months and ISO 42001 in 93 days.
Scott didn’t take the credit for the paperwork. He took it for the years before.
“I see the certifications as recognition of work that’s been going on for a very long time.”
Scott Dietzen, at 9:16
The company came together, he said, with a strong AI research team from places like Google, Meta and NVIDIA and “an extremely strong systems team, people that helped design Snowflake, that helped design Databricks.” Customers were trusting Augment with their software, so it needed “a very mature attitude about what it was going to take to build a highly secured cloud service” from the start. Being first, in his words, “with each of those proof points, attestation points along the way,” was a sign of how security was designed in rather than added. It is the same lesson episode 77 with Augment’s founding engineer Dirk Meister tells from the inside, and the reason the audits were quick: the controls were already in the system.
No way for intellectual property to leak
Sasha asked the question every buyer of these tools asks first.
“What is the most common question that you hear from the prospects and customers around data privacy, especially and explicitly their source code?”
Sasha Sinkevich, at 10:18
Customers want state-of-the-art protection, Scott said, and something simpler under it.
“They want to understand that there’s no way for intellectual property to leak.”
Scott Dietzen, at 10:26
He drew the ethical line in two steps. Defaulting a customer into sharing code for training at least implies the vendor uses it internally. Turning around and selling that code to a third party for training is the step he called “a very unethical behavior,” and he said Augment had heard reports from model providers that some competitors were doing it. We couldn’t verify that, so we’ll leave the who out. His rule needs no names: never take on any behavior “where somebody’s proprietary IP is crossing a boundary into another organization without their knowledge.”
Four days after his post on agents, he turned the rule into four questions anyone can ask a vendor, in a May 2025 post titled “4 Essential data questions for your AI coding platform” (Augment Code). Augment posted them as a thread the same day.
1. Have you ever trained models on our code without
Augment Code (@augmentcode) · May 6, 2025
explicit permission?
2. Is training disabled by default?
3. Has our data ever been sold to third parties?
4. Can we enforce org-level restrictions, even if
someone opts in by mistake?
The post publishes Augment’s own answers, no, yes, no and yes, and then gives the advice I’d give any founder, “We strongly recommend getting these answers in writing,” before any AI indexes your code. Behind the questions it lists what a buyer should expect to see, a contract that rules out training and retention, SOC 2 and ISO 27001 audits, indemnification for your IP and proof of possession, and it borrows an old idea about software, core and context. Core is the code with “unique, material value,” the trade secrets and patented algorithms. Context is everything around it. A company might happily trade context code for a better model, the post says, but in practice there’s no clean way to separate the two, so the safe default is to protect it all, and to try a new tool on lower-risk code bases first.
The scorecard further down turns those questions and four more into a score for the assistant your team uses today.
Data is the new gold standard
Sasha had the one-line version of why vendors want the code.
“At this point, it’s pretty clear that data is the new oil. All of the models that are available on the market have been trained on all of the publicly available information. And at this point, data is the new gold standard.”
Sasha Sinkevich, at 11:20
Scott agreed and went to the question underneath it.
“Yeah, I think the open question is, are models going to start to plateau because we’ve tapped all of the publicly available software data in the world?”
Scott Dietzen, at 11:35
There weren’t real algorithmic or compute advantages between the large model providers, in his view, so data was the difference. Synthetic data could help, but only with “really strong evaluation functions,” and software doesn’t have one built in. His comparison was Go: the space of program designs is like the space of Go strategies, “But with Go, there’s an outcome that you see, whether you win or lost the game. We don’t necessarily have that with software.” He expected the progress to continue “through next year” and wondered whether it would flatten after that.
He had made the same bet in public that February, in six predictions for 2025. The first was that the moat around large language models would shrink and value would move to applications. The second was that enterprises would invest in specialized models as the general ones level off, “as the curve inevitably flattens” (Augment Code). For a security team the practical reading is the same either way. If data is the gold, your code is part of the vault, and the four questions above are the lock.
Don’t disrupt the developer
Sasha asked what matters most for winning over the people who use the tool, the developers with “a lot of pain points” and “a lot of tickets.” Scott had two answers. The first was about where the product lives.
“One, I think it starts with, don’t disrupt the developer unduly. So we have structured our product as a plug-in to their existing development environment.”
Scott Dietzen, at 13:03
If a developer is on VS Code “in the Microsoft ecosystem with the Microsoft marketplace and all the Microsoft security and support,” Augment keeps all of that. The second answer was context again, with the sharpest version of the anti-vibe coding line.
“Without context, it’s all you can do is vibe code. With context, then the engineer has a true partner.”
Scott Dietzen, at 13:03
An AI with context “may understand the code base better than that engineer does,” he said, and that is the difference between an AI for software engineering and the tools that only cut it for new projects started from scratch.
Meeting developers where they work goes past the editor, too. Scott’s January 2025 post argued that software development happens beyond the IDE and described engineers pulling Augment into a Slack thread about an error message, where it answered from the code base and suggested the fix (Augment Code).
It basically puts you in the IDE business
The plug-in choice had a security story, and I couldn’t resist telling mine. When I first saw other AI coding platforms fork VS Code, one word came to mind.
“I think the word that came to mind when I saw that was, whoops.”
Jon McLachlan, at 14:01
A fork starts easy, I said, and then it diverges, and keeping up with Microsoft’s changes and your own engineers’ at the same time “just becomes more and more challenging as time goes forward.” Scott put it in one sentence.
“Agreed. I mean, it basically puts you in the IDE business.”
Scott Dietzen, at 14:33
The team behind VS Code, he said, “is not small,” larger than most AI coding startups’ entire staffs, and keeping a fork current with what Microsoft and JetBrains ship is “going to be very painful over time.” He had argued the whole case in March 2025, in a post titled “To fork or not to fork?” (Augment Code). GitHub Copilot and Augment are plug-ins inside the standard IDEs, VS Code, the JetBrains family and Vim, while two of Augment’s best-known competitors had modified VS Code’s source instead. A fork, in his list, costs its users a switch of IDE, Microsoft’s support, the plug-ins Microsoft licenses and certifies, the shared marketplace and timely access to new VS Code features. He gave the forks their due too, a head start in the market while Augment perfected a harder-to-build plug-in, and noted with a wink that one of those forked editors was among Augment’s most popular places to try the plug-in.
Sasha added the attacker’s view, which is the one I’d keep.
“When you fork something and you cannot keep up with the maintenance of the project, there is an exposure.”
Sasha Sinkevich, at 14:59
AI tools help red teams and would-be red teams find vulnerabilities much faster, he said, so a project that falls behind its upstream invites “a little bit more discovery activities by members of society.” Scott closed it with the reason most enterprises pick an ecosystem in the first place: “Microsoft has a large security team that does due diligence on their software, and that’s why people opt into the Microsoft ecosystem.”
The useful question for your own team is not which tool to buy. It’s how your editor gets its security fixes, and how quickly. If the answer is whenever the vendor gets around to it, that belongs in your risk register next to the four data questions.
It needs to be easy to adopt
I thanked him for taking security seriously, because the thing I care about most as an entrepreneur is getting new technology into people’s hands, and “it needs to be easy to adopt.” He had learned that the hard way.
“I think all new tech has got to be easy to adopt. This is one of the lessons I learned early in my career, with some software that was very hard to use and wasn’t necessarily robust.”
Scott Dietzen, at 16:27
He came out of graduate school “very naive about how easy and reliable commercial software had to be,” and he has carried the lesson to every company since: “it’s got to be unbelievably easy, or you’ll never get the mass adoption that we startups aspire to.” Security is part of easy. A control nobody can live with gets turned off, and a plug-in that keeps the editor a developer already trusts is security a team will actually keep.
Do you already have that data?
There was one design I wanted listeners to hear about, because it could only have been built in on day zero. Augment’s retrieval asks a different question before it shows anyone a piece of code.
“The answer to the question, are you allowed to see this piece of data, is actually driven by the answer to the question, do you already have that data?”
Jon McLachlan, at 16:57
If you already have the file, of course you’re authorized to see it. That is proof of possession, and it “solves tons of engineering questions,” like how to deduplicate what the whole team shares, while keeping what only one engineer has out of everyone else’s results. Scott named the risk it answers, and it’s one most buyers haven’t thought about.
“And people don’t realize that AI can be a way for intellectual property to leak. And so if the AI I’m working with is informed by code that I’m not allowed to see, all of a sudden I get insight into algorithms that I’m not supposed to be privy to.”
Scott Dietzen, at 17:51
The design is written up in public. Augment’s January 2025 post on its real-time index, by Dirk Meister and Markus Rabe, has the editor prove it holds a file by sending a cryptographic hash of its contents (Augment Code), and Scott’s May post lists proof of possession among the basics, “so the AI cannot inadvertently share your most secure code.” Dirk walks through it on episode 77, including why a stolen token reaches the service but not the code. For any assistant you evaluate, the question to ask is the same one: when it retrieves code to answer someone, does it check that this person could already see that code?
Going well faster than that
Sasha brought a prediction to test.
“That’s right. There is an interesting study that predicts that about 50% of all of the enterprise software engineers will use ML-powered coding tools by 2027.”
Sasha Sinkevich, at 18:18
The study is a Gartner prediction, and Augment’s own launch release in April 2024 quoted it: “by 2027, 50% of enterprise software engineers will use ML-powered coding tools” (Augment Code). Gartner had a newer forecast by then. On April 11, 2024 it said that by 2028, 75 percent of enterprise software engineers would use AI code assistants, up from less than 10 percent in early 2023, and that 63 percent of organizations were already piloting, deploying or running them (Gartner). Scott’s answer was shorter.
“I think it’s going well faster than that.”
Scott Dietzen, at 18:41
A survey from the same year backs him. In Stack Overflow’s 2025 Developer Survey, with more than 49,000 responses from 177 countries, 84 percent of respondents were using or planning to use AI tools in their development process, up from 76 percent the year before, and 51 percent of professional developers used them every day (Stack Overflow). The same survey has the reason the tech lead’s review still matters: the most experienced developers were the most cautious, with the lowest rate of high trust in AI output, 2.6 percent, and the highest rate of high distrust, 20 percent.
His own numbers are from Augment’s customers and are his account: 40, 50 and 60 percent productivity gains in the early days, which he dated to about nine months before we recorded, and “two, three X kinds of improvements” with agents, with the best developers getting the most out of it. The public record has one customer’s number from the launch, a chief executive quoted at “a >40% increase in developer productivity across the board” (Augment Code). The part of his answer I’d frame is where the gains go. Every piece of software has a wish list of architecture work, debt, features and security fixes.
“What if we could pay that all down? What if we could deliver, you know, all the software we aspire to?”
Scott Dietzen, at 18:41
The security fixes are on that list. A team that pays its backlog down with agents can finally get to the vulnerability tickets that have been waiting since the last penetration test.
A first seven-figure deal
I asked for his proudest day so far at Augment. He couldn’t pick one.
“I can’t point to one single day. There’s been a bunch of customer victories. Our first six-figure deal, our first seven-figure deal, those are natural milestones.”
Scott Dietzen, at 20:14
The one he chose was self-service: “getting the product to the point where we were comfortable that every developer that wanted to try it in the world would have a good experience.” That is a quality bar before it is a sales channel. Augment Code launched for teams on October 24, 2024 (Augment Code), the “nine months ago” he kept coming back to, and by April 2025 anyone could start a free trial of its agent (Augment Code). It also fits the lesson from his grad school days. Software that anyone can try has to work for anyone who tries it.
Agents that wake up
Then I asked what success looks like, as far out as he wanted to look. The first answer was the market he’d always named: to be “the leading product for real software engineers,” the people looking after “the tens of millions of lines existing production code bases.” The second was bigger, and he was careful about one word in it.
“You know, we see agents as being able to help automate the full software development lifecycle, not unassisted, right? So our vision is that it is the combination of human intellect and agent, machine intellect.”
Scott Dietzen, at 21:02
Developers do a lot more than write code, so the agents have to help “all the way along the journey.” Then he described the agents I think most teams will meet first, the ones that don’t wait to be asked.
“Like, there’ll be agents that wake up in the event of a site failure or a break in [a] CI/CD pipeline. Agents can be, or a security attack, right? The agents can wake up and start doing work”
Scott Dietzen, at 21:02
They report back on what happened and what they recommend. Augment had started down that road that spring. Its Remote Agent, announced on May 7, 2025, ran the agent in the cloud to work through “we’ll-get-to-it” tickets in parallel (Augment Code), and on June 5 it was generally available in VS Code, letting engineers “run multiple development tasks concurrently in the cloud” (Augment Code). The security version of his picture is worth writing down before it arrives. An agent that wakes up during an attack needs to know what it may touch, who it reports to and when it must stop and ask, and those answers belong to people, written before the night it wakes up.
I don’t think the next big leap in software productivity comes from “vibe coding.” It comes from removing the grind in real codebases.
Eric Schmidt (@ericschmidt) · May 7, 2025
@AugmentCode’s new Remote Agent tackles flaky tests, stale docs & tedious refactors—with up to 10 autonomous agents. Learn more: https://t.co/ESPfzLymO1
He also put a number on how fast the agents were growing up.
“Six months ago, you would struggle to have an agent do something that took you 10, 15 minutes. Now I think an agent could take on an hour, 90-minute kind of task.”
Scott Dietzen, at 21:02
Next year, he hoped, “that doubles and doubles again.” That is faster than the best public measurement, though not by much on recent models. METR, the research group that measures how long a task AI agents can finish, found in March 2025 that the length of task frontier agents complete with 50 percent reliability had been doubling about every seven months for six years, and faster than that when it fit only the 2024 and 2025 models (METR). Fifteen minutes to an hour or more in six months is two doublings, which is his account of Augment’s agents rather than METR’s average. The calculator further down runs the same curve on your numbers.
The last part of his answer is the one I’d send to every board that asks how many engineers AI will let them cut.
“These businesses that use this as sort of a private equity [play] to cut their teams and do the same work with fewer are missing a colossal opportunity”
Scott Dietzen, at 21:02
Every software-dependent business he knows could grow substantially with better software, and “the software abundance opportunity” is what he called the investment thesis. He’d made the same prediction in February, that “Coding AIs increase demand for software engineers” (Augment Code), and in January he argued that once AI makes software investment predictable, companies will spend more on it, not less (Augment Code).
The context costs the square of its length
Sasha came back to the context engine, and how much the quality of what goes in decides what comes out. Scott explained why Augment built one instead of taking either of the shortcuts of the time.
“Yes, you know, there were naive approaches where some people thought they would fine-tune these models on company software.”
Scott Dietzen, at 24:02
Fine-tuning a model on a company’s code has two problems in his telling. People get nervous about leaking IP, and, “much more so,” the code changes constantly while training is expensive. Nobody is going to train a new model every time someone pulls a branch, and “You want your AI to be current.” The other shortcut was to pass the whole code base to a model with a very long context window.
“You know, these models, of course, the context costs the square of its length. And so any notion of passing in that much context is ludicrous. It will be horrifically slow.”
Scott Dietzen, at 24:02
The square is the attention mechanism inside a transformer, which compares every token with every other, so doubling the context roughly quadruples the work. His human analogy is the better argument. If you had read a hundred-million-line code base and then got asked random questions about it, “you would have to refer back,” and so do the models. Hence retrieval: find the right few pieces of the code for each question and pass those. He had predicted in February that retrieval would beat fine-tuning (Augment Code), and in June Augment published how it kept retrieval fast on code bases of more than 100 million lines, an approximate search that keeps “more than 99.9% fidelity” to the exact results and made real-time features like completions more than 40 percent faster on its hardest cases (Augment Code).
Retrieval is also where the security lives. Whatever decides which pieces of code reach the model also decides what one user’s question can expose, which is why the possession check from earlier sits in the same pipe.
Metaprogramming
I asked what stands between here and his future. “Truthfully, there’s a lot of developer education that still needs to happen,” he said. The most forward-leaning developers were having huge success, and the rest of the market hadn’t seen yet how capable the tools were. The biggest change was in how the best of them worked.
“Many of the developers on Augment have moved to what I call metaprogramming, where they don’t necessarily directly manipulate the code themselves anymore. They count on the AI to do it for them.”
Scott Dietzen, at 25:34
They spec out to the AI what they want, let it generate the code, look at it and tweak it, all through the AI. It is a different paradigm, and some developers have to watch a colleague do it before it clicks, which is where he reached for an old line.
“And so there’s that old quote: the future is already here, but it’s not evenly distributed.”
Scott Dietzen, at 25:34
The quote is usually credited to the science fiction writer William Gibson, and Quote Investigator traces the earliest full version in print to a 1992 newspaper article that attributed it to him (Quote Investigator). Scott had written the same split into his February predictions, that AI-augmented engineers “will toggle between programming and meta-programming” (Augment Code).
Sasha had the picture for it: “So instead of molding and building a Lego piece, you can actually build a whole model out of many Legos.” Scott made it a ship.
“Yes. And do it in parallel, right? You can have one agent work on, whatever, the hull to your ocean liner, and others could start building the cabin and the engine and the propellers.”
Scott Dietzen, at 27:02
Seeing how the pieces fit together is still human insight, he said. Programming is tedious because computers need an extremely precise description of what you want, and most programs have no specification apart from themselves: “The spec is the software itself.” An AI that can read that specification and help evolve it is “incredibly powerful.” For security, the spec point is the one to keep. If the code is the only specification, then the security requirements had better be written somewhere an agent will read them, in the repository, next to the code.
The biggest productivity unlock
Sasha put it in the history of programming. Everyone used to work at the level of the compiler, then Java promised one language that could run everywhere, and now we’re “moving up another stack, where you can speak the human language.” Scott has been around for most of that climb.
“Yes, I’ve been around long enough that I’ve, you know, I wrote some BLISS and some machine code early on.”
Scott Dietzen, at 28:19
BLISS is a systems programming language built at Carnegie Mellon around 1970, by W. A. Wulf, D. B. Russell and A. N. Habermann, and perhaps the best-known systems language until C took over (Wikipedia). He’s a believer in the third- and fourth-generation languages, and he ranked the newest step above all of them.
“But I do think AI is the biggest productivity unlock that we will see in this industry, you know, throughout its entire history. And it may be the last one that we need in terms of improving how we get software into the world.”
Scott Dietzen, at 28:19
He made the comparison in writing in August 2024, when he predicted that the next decade would see the biggest leap in software quality and team productivity “since the advent of high-level languages” (Augment Code).
Testing was what users were for
Students ask me what to study now, so I asked him for advice for young engineers facing all this change early in their careers. He started with a belief.
“I am still a big believer in human intelligence and the role that human insight has in complementing machine intelligence.”
Scott Dietzen, at 29:23
He gave a talk at his alma mater about 12 years before, he said, when the big worry was that all the interesting work in computer science had been done, “And I just found that, like, to be the most preposterous statement.” Then he told the story of his own career as a lesson in how hard software is.
“Augment is the fifth startup that I’ve gotten to be part of. All of them depended on large-scale software. And the first four, horrific pain.”
Scott Dietzen, at 29:23
He hadn’t understood, before the real world, how hard it is to make software “reliable enough, easy to use enough, secure enough,” and he named the reason.
“You know, I had come out of a grad school where testing was what users were for, not something that we had to do in advance.”
Scott Dietzen, at 29:23
The rest of his path is in his own August 2024 post. He spent “a decade at CMU collecting degrees,” with Geoffrey Hinton teaching his introductory AI class and a front row seat for the old debates between symbolic reasoning and neural networks, and wrote “a dissertation on a programming language theoretic approach to machine learning” (Augment Code). On the tape he called it “a thesis in machine learning, albeit a long time ago,” and said the fun of Augment is seeing that thread and large-scale software finally meet. After Pure Storage he wasn’t sure he’d work again. Augment’s early team invited him into their research sessions while he was “on the beach,” as he put it in the post, and he joined in early 2023.
“I had a major fear of missing out in terms of these two threads, machine learning and large-scale software, coming together.”
Scott Dietzen, at 29:23
For a student, I’d take two things from that. The interesting work is never done, and the discipline that will matter most is the one he learned last, making software reliable and secure before a user finds out it isn’t.
Trust your instincts
I asked what he’d tell his younger self. The first lesson was the one from his first startup.
“At the first startup, I learned how hard it is to make software commercially viable. You know, to get mass adoption, software has to be really easy, and it has to not break and do things wrong.”
Scott Dietzen, at 31:12
“Ironically, as software ages, it can grow a lot of hair, right?” His example was the Oracle database, painful to adopt today and “incredibly easy to build and deploy” in its early days. Another was Linux.
“I remember sharing with my Windows friends that we had Linux servers that had been up for years and never been rebooted, and they were just like, how is that possible?”
Scott Dietzen, at 31:12
The second lesson was harder to learn. That first startup “went sideways for a while,” with customers but no inflection, and five years in he told his mentor he could see two endings and had missed the third.
“I knew we could hit the moon, I knew we could crash and burn, but I didn’t know we’d be somewhere in between.”
Scott Dietzen, at 31:12
If he had trusted his instincts, he’d have seen sooner that they were missing the target. “And so, trust your instincts. Of course, hone them through hard work and good luck.”
Radical transparency
Sasha asked which lessons he carried into an AI startup, for the team, the culture and the market. “Entrepreneurship is a team sport,” Scott said, and his first rule follows from it: check your ego at the door and avoid politics, because “politics kills organizations.” Keep the team looking outward, at serving customers and winning business, and trust that the pie will be fairly shared. Then two practices I haven’t heard a chief executive describe this plainly.
The first is about pay. At his companies, he said, people are told they don’t need to negotiate their own offers, because a leadership team that works hard to get rewards right shouldn’t reward negotiation. He has heard managers at larger Silicon Valley companies tell an engineer to bring a competing offer before they can give a deserved raise.
“That’s a ludicrous behavior to me, because you should reward people for the quality of their work, not the fact that they could potentially defect somewhere else.”
Scott Dietzen, at 33:02
The second is about information.
“Another thing I like is radical transparency.”
Scott Dietzen, at 33:02
When information passes selectively down the tree, it becomes power that can be used politically. So broadcast as much as you can.
“One of the things we do at Augment is we share our board decks with the entire company, as well as the results of the conversations, because we want everyone to have insight into not just what’s going well, but what we need to work on.”
Scott Dietzen, at 33:02
A chief executive who only cheerleads, he said, misses “a huge, huge opportunity.” Sasha connected it to our side of the work: “It’s the internal trust, but internal trust, it transitions into the external trust that is established between the company and customers.” That is the most useful sentence in the episode for a security program. A company that shares bad news inside will report an incident honestly outside, and buyers can tell. Scott agreed, and added the weight of it at Augment’s scale, a challenge “when you’ve got hundreds of thousands of developers using your product all around the world.”
Stone soup
I had a confession to make. I joined Pure Storage after Scott had moved on, “but it stuck,” the culture he left, and it spoils people because it’s so good. He gave the credit away with a folk tale.
“It was very much a team effort, right? I used the stone soup analogy. I didn’t build the Pure culture or define it myself. But I did bring the pot, and I talked about the importance of culture”
Scott Dietzen, at 36:32
In the story, hungry strangers get a whole town to add a little of what it has to a pot of water with a stone in it, and everyone eats (Wikipedia). At Pure, by his telling, everybody showed up and chipped in, “and, you know, we kept the best and shed the rest and ended up with a really good culture.” It is a good model for a security culture too. Nobody can impose one. Someone brings the pot and keeps talking about why it matters, and the team fills it.
We ended on thanks in both directions. “I could not be more thrilled to have YSecurity on this journey with Augment,” he said, which is the disclosure from the top of the page in his words.
What happened after the tape
His LinkedIn has the rest. In June 2025, a few weeks before this episode aired, he joined the board of Redis, part time. His time as Augment’s chief executive ended in July 2025, the month the episode went out, and he has stayed on Augment’s board, alongside seats he already held at HYPHY, as vice chairman of Pure Storage and on the dean’s advisory board of Carnegie Mellon’s School of Computer Science (LinkedIn). His seat on the board of Sigma, the analytics company, ran from March 2018 to May 2026, the month he posted that it had closed an $80 million Series E at a $3 billion valuation (LinkedIn). Pure Storage itself took a new name in February 2026 and is Everpure now (Wikipedia), still with Charles Giancarlo, the chief executive Scott recruited to succeed him in 2017, at its head.
For Augment, the certification he called a proof point became a public one too. On August 5, 2025, a week after the episode, Coalfire announced Augment’s ISO/IEC 42001 certification as a first for an AI coding assistant (Coalfire). And on September 10, 2025, Augment’s founding engineer Dirk Meister came on episode 77 to walk through the architecture Scott keeps pointing at, proof of possession, customer managed keys and access that opens only with a reason. The two episodes are a pair. Scott’s is why the code base is the hard part. Dirk’s is how you keep it safe while an AI reads it, and episode 43, with Evan Driscoll and Dion Almaer in 2024, is where the story starts.
The phrase from this episode I hear founders use most is his first one. A developer as a tech lead for a team of agents is a job with security built into its description: hand out the work, keep the design, and check every result before it ships. The teams that write those checks down now, along with what each agent may touch, are the ones whose buyers will trust their agents later.
Did you catch it? A six-question quiz
Interactive · six questions
Did you catch it?
Six questions on what Scott, Sasha and I actually said, with the timestamp for each answer. The answers are under each question for anyone reading without JavaScript.
-
How does Scott describe Augment in one line?
- The copilot for every developer
- The anti-vibe coding company
- The IDE built for agents
b. "You can think of us as the anti-vibe coding company, if you like." Transcript, 0:27.
-
What does Scott say the developer's role will become?
- A reviewer of pull requests written by AI
- A prompt engineer
- A tech lead for a team of agents
c. "the role of the developer is going to move to being a tech lead for a team of agents." Transcript, 2:37.
-
Forking VS Code, in Scott's words, basically puts you in which business?
- The IDE business
- The marketplace business
- The security patch business
a. "Agreed. I mean, it basically puts you in the IDE business." Transcript, 14:33.
-
With proof of possession, the question "are you allowed to see this piece of data?" is answered by which question?
- Is your manager on the access list?
- Do you already have that data?
- Did you sign in from a trusted device?
b. "is actually driven by the answer to the question, do you already have that data?" Transcript, 16:57.
-
What does Augment share with the entire company, by Scott's account?
- Every customer contract
- Each engineer's salary
- Its board decks, and the results of the conversations
c. "One of the things we do at Augment is we share our board decks with the entire company, as well as the results of the conversations" Transcript, 33:02.
-
Which number startup is Augment for Scott?
- The fifth
- The third
- The second
a. "Augment is the fifth startup that I've gotten to be part of." Transcript, 29:23.
Is your AI coding assistant ready for the code that matters?
Scott’s four data questions are the start of an evaluation, not the end of it. This scorecard asks eight questions about the AI coding assistant your team uses today, from what it may train on to how your editor gets its security fixes and what an agent may do without asking.
Scorecard · eight questions
Is your AI coding assistant ready for the code that matters?
Answer for the assistant or agent your engineers use most, on your most sensitive repository. If two answers fit, pick the lower one.
-
Has the vendor told you whether it trains models on your code?
- We haven't asked
- Yes, on its website or in a sales call
- Yes, in our contract, with training and retention ruled out
Scott's first question, and his advice: get the answers in writing. No way for intellectual property to leak has the four.
-
If one engineer opts in to sharing code for training, what happens?
- Our code is shared from then on
- Training is off by default, but each person can turn it on
- An admin setting blocks it for the whole organization
The fourth question in Augment's thread, and the one most teams forget.
-
Has your code, or data about it, ever gone to a third party for training?
- We don't know
- The vendor says no
- The contract says no, and lists every subprocessor that sees our code
The step Scott called "a very unethical behavior" is the one you can rule out on paper.
-
When the assistant retrieves code to answer someone, does it check that this person could already see that code?
- No, or we can't tell
- It respects repository permissions
- Yes, with proof of possession or an equal check, file by file
Otherwise the AI can share "algorithms that I'm not supposed to be privy to." See Do you already have that data?
-
How does the editor your team uses with it get security fixes?
- It's a fork, and we don't know how far behind upstream it runs
- It's a fork that tracks upstream on a published schedule
- It's a plug-in in a standard editor we already patch
Sasha's point: "When you fork something and you cannot keep up with the maintenance of the project, there is an exposure."
-
What can its agent do without asking a person first?
- Anything the developer can
- Edit files, but commands and integrations need approval
- Only what we wrote down for each task, with approval for commands, snapshots to roll back and least-privilege credentials
Augment's own agent shipped with a manual mode that pauses for approval, and snapshots for one-click rollback.
-
What evidence have you seen from the vendor?
- A security page
- A SOC 2 Type 2 report
- A SOC 2 Type 2 report, plus ISO 27001 or ISO 42001 and indemnification for our IP
Scott's May 2025 post lists the same evidence. Read the report's scope, not just its cover.
-
Where did you start using it?
- Everywhere at once, including our core code
- On core code, with a pilot team
- On lower-risk code bases first, then the core once the answers were in writing
Core is the code with "unique, material value," in Scott's post. Start everywhere else.
The assistant may be excellent, and you don't know yet what it does with your code. Start with the paper: send the four questions to your vendor, put the answers in the contract, and turn training off at the organization level. Then check whether retrieval respects who can see what, and keep it on lower-risk code until it does.
The data questions are mostly answered. The next layer is the agent: write down what it may do for each kind of task, require approval for commands and integrations, give it credentials scoped to the task, and find out how quickly your editor gets upstream security fixes.
Your contract, your settings and your vendor's architecture all keep the code where it belongs, and your agents have written limits. Now test it: ask the assistant, as a user without access, about code that user shouldn't see, and write down what came back.
How long a task can your agents take on next year?
Scott’s estimate was that agents went from 10 or 15 minute tasks to an hour or 90 minutes in six months, and he hoped for “doubles and doubles again.” METR measured a doubling about every seven months. This calculator projects the curve forward from where your agents are today, and shows how much review each task will need.
Calculator
How long a task can your agents take on next year?
The longest task, in minutes of a person's work, that your agents finish reliably today, how often that doubles, how far ahead to look, and how many minutes of review a person spends per hour of agent work.
If your agents finish 60-minute tasks today and that doubles every 7 months, in 12 months one task is about 3.3 hours of work, 3.3 times longer, and needs about 33 minutes of a person's review at 10 minutes per agent hour.
The seven-month default is METR's March 2025 measurement for frontier agents at 50 percent reliability, which was faster on the 2024 and 2025 models alone. The 60-minute default is the low end of Scott's estimate on the tape, and the review rate is a placeholder for yours. A trend line is not a promise, and longer tasks are where the tech lead's checks matter most.
From BLISS to a team of agents
Everything on this page sits on one timeline, from a systems language at Carnegie Mellon to agents that wake up on their own. Filter it by thread.
Timeline · 1970 to 2026
From BLISS to a team of agents
Filter by thread. Dates are from the public record except where marked as Scott's own account on the tape.
- BLISS
A systems programming language built at Carnegie Mellon, and one Scott wrote early on (Wikipedia).
- "The future has arrived"
The earliest full version in print of the line credited to William Gibson (Quote Investigator).
- Zimbra
President and chief technology officer (LinkedIn).
- Yahoo buys Zimbra
And Scott runs applications at Yahoo until October 2009 (Wikipedia).
- Pure Storage
Chief executive until August 2017, from 15 employees to nearly 2,000 (Augment Code).
- The IPO
Pure Storage lists on the New York Stock Exchange (Wikipedia).
- Chairman
Charles Giancarlo becomes chief executive, and Scott chairs the board, then becomes vice chairman in 2018 (LinkedIn).
- $2.41 trillion
The cost of poor software quality in the U.S., by CISQ's estimate (CISQ).
- Augment is founded
By Igor Ostrovsky and Guy Gur-Ari (Augment Code).
- Chief executive
Scott joins Augment, after sitting in on its research sessions (LinkedIn).
- 75 percent by 2028
Gartner's forecast for AI code assistants, up from less than 10 percent in early 2023 (Gartner).
- Out of stealth
A $227 million Series B at a $977 million valuation, $252 million in all (Augment Code).
- Evan Driscoll and Dion Almaer on this show
Augment's first episode with us (episode 43).
- SOC 2 Type II
After an audit window from April 1 to July 1 (Augment Code).
- A personal journey
Hinton's intro to AI class, a decade at CMU, and an AI that deletes code (Augment Code).
- Augment Code launches
Developer AI for teams, the "nine months ago" on the tape (Augment Code).
- No such thing as a simple change
An Airbus A380, remodeled in flight (Augment Code).
- Six predictions
Retrieval over fine-tuning, metaprogramming and more demand for engineers (Augment Code).
- Next Edit
The ripple effects of a change, across the repository (Augment Code).
- To fork or not to fork?
The case for a plug-in (Augment Code).
- A seven-month doubling
The length of task agents finish, measured over six years (METR).
- 65.4 percent
On SWE-bench Verified, the top published open-source score at the time (Augment Code).
- Augment Agent
From a new app to a monorepo with more than 100,000 files (Augment Code).
- Four data questions
What to ask any AI coding vendor, in writing (Augment Code).
- ISO/IEC 42001
Certified by Coalfire, with YSecurity running the program (Augment Code).
- Redis
A board seat, part time (LinkedIn).
- Remote Agents
Generally available in VS Code, many tasks at once in the cloud (Augment Code).
- 100 million lines
Retrieval more than 40 percent faster on the largest code bases (Augment Code).
- 84 percent
Of developers using or planning to use AI tools, and 51 percent of professionals using them daily (Stack Overflow).
- To the board
Scott's time as chief executive ends, and he stays on Augment's board (LinkedIn).
- Episode 74 airs
This conversation. "a tech lead for a team of agents."
- A first, by Coalfire's headline
The ISO/IEC 42001 certification, announced (Coalfire).
- Dirk Meister on this show
Augment's founding engineer, on why security starts at line one (episode 77).
- Jack Cable on this show
Corridor, on what gets missed when nobody reviews the code (episode 84).
- Everpure
Pure Storage takes a new name (Wikipedia).
- Sigma's Series E
$80 million at a $3 billion valuation, as Scott's eight years on its board end (LinkedIn).
What to do this week
Five things, in the order Scott would ask for them.
- Send the four questions. Ask the vendor of every AI coding tool your engineers use whether it has trained on your code without permission, whether training is off by default, whether your data has ever gone to a third party, and whether you can block opt-ins for the whole organization. Get the answers into the contract.
- Turn training off at the top. Find the admin setting that keeps any one engineer from opting your code base in, and turn it on. Then check who else can change it.
- Test what retrieval shows. Sign in as someone without access to your most sensitive repository and ask the assistant about it. Whatever comes back is what proof of possession, or its absence, means for you.
- Find out how your editor gets its fixes. If your team uses a forked editor, learn how far behind its upstream it runs and how fast security fixes arrive, and put the answer in your risk register.
- Write the agent’s job description. For each agent, write down what it may touch, when it must stop and ask, who it reports to and how to roll back its work. Start it on incremental, well-defined tasks in lower-risk code, the way Scott recommends, and widen the scope as it earns it.