TS

Tarun Sachdeva

Things Tarun Says on Podcasts

Where to Find Them

Tarun Sachdeva has been a guest on devtools.fm: Developer Tools, Open Source, Software Development .

Recently: “Tarun Sachdeva - Traces” on devtools.fm: Developer Tools, Open Source, Software Development (November 2026).

What They Said

“There's actually nothing in Traces data that companies don't already have, all the way from your Slack history to any of these other things. If someone is SSHing into a production box they shouldn't be, that data can already be acquired through the MDM tools and all of the other ways that companies are monitoring people's behavior on their devices. But Traces as a product is a workflow tool, a tool for learning, a tool for teams.” — Tarun Sachdeva, devtools.fm: Developer Tools, Open Source, Software Development

Asked whether recording agent sessions could make engineering feel like surveillance or push token-maxing, Tarun, who says engineering leaders ask him this daily, argues the data isn't new and how a leader uses it is a choice.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-02 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 43:45 into the episode
Tarun Sachdeva

…in some sense, but like it's so different in every harness and every model. It's just like, how useful is that number really? Is like attaching the number of bytes your code base like consumes. It's like, okay, cool, useful in some contexts, but for the most part, it's just not useful for workflow.

Andrew Lisowski

…coding because we're all using agents nowadays. And seeing that thought process behind that would be huge because everybody's accusing everybody else of slop right now. But I don't think you'd accuse them of slop if you saw their month-long chat with AI that like probed the depths of mathematics.

Tarun Sachdeva

Yeah, exactly. And those conversations are some of the most interesting. You know, we have some really interesting ones on our that are shared publicly as well. And I think that, you know, your point here is exactly very astute is that agents are the primary interface to really all knowledge work. And obviously software development is where that's sort of led the way there. And there's an argument that a lot of knowledge work can be encoded as software development, but so much knowledge work, whether that's sales, marketing, you know, any kind of operational role, people are using agents for everything. And every single time they use these agents, they are, this is not just like a log of your computer's activity. This is decision data. This is your decisions, your team's decisions. And that decision, that level of fidelity into those decisions should benefit the teams that are using that whose data it is, right? Not just the model providers that are providing the models, right? So a lot of this is like, you know, all about that. That like this idea of like these agents becoming useful beyond development is something that we're seeing as well. We're seeing a lot of sales team useTraces.com in surprising ways on…

Justin Bennett

There's something that I would, I'd love to sort of like chat through. As you were talking about like thinking of like these certain features as like a call center or something, there's like, you know, classic examples of like engineering orgs getting treated as call centers and that like really…

Tarun Sachdeva

Yeah, yeah. So great question. And very much a question I talk to engineering leaders every day. And the one question that they ask is like, what happens when my team says this feels like surveillance? Right. Or what happens when my team feels like, hey, this is data that they shouldn't, that is…

Andrew Lisowski

well, that wraps it up for our questions. Thanks for coming on and talking about this. This is a super interesting dive into really what the future of coding and knowledge work could look like. So thanks for coming on and talking about it.

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“If the volume of PRs increases, then the amount of time that we spend on any given PR needs to decrease. And in order to maintain the quality of review without just saying 'forget it, let's just hit accept', we need better tools to get to an understanding of what is happening in this PR faster… The process of development is a session, but the process of review is also a session with an agent, where an agent can help you understand what is happening inside a PR and help you probe and ask questions. A code review is another form of conversation.” — Tarun Sachdeva, devtools.fm: Developer Tools, Open Source, Software Development

Asked how his own team reviews code now, Tarun admits they haven't solved it: they have dozens of open PRs and still read the code. He sets out why he thinks the pull request needs rethinking for the agent era.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-02 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 20:43 into the episode
Speaker 4

…they land. So if your team is tired of bloated project management, or maybe you're just like starting fresh and you just want one tool instead of many, give macro a try. It's fast, it's fun, it's a better way to build. Sign up at macro.com and get $100 off your subscription. Using DevTools 100.

Justin Bennett

So I'm curious, when you're actually performing a review, are you just using an agent within MCP to probe different parts of the code? What is your process? How's your process in doing a review on GitHub or whatever changed?

Tarun Sachdeva

Man, I don't want to say that we figured it out. I think that I think pull requests and just code review is just probably the biggest pain point for every development team. I don't want to sit here and pretend like we've gotten that stuff solved. As I'm talking, we're sitting on 40 open PRs and there's a lot of work that we have to do to figure out how PRs are going to really work for us. Because one thing that we care about, we still care about is we do still read the code. So we try to understand what's happening. We try to make sure that we can reason around these things. And perhaps that's a kind of relic of the past in some ways. And there's arguments for and against that. And it's an interesting conversation, of course. And very much, we're not trying to hold on to something that is maybe not relevant. But right now, for us, it is quite relevant. We're building product that requires us to understand what is happening at every step of the way. Andrew Lisowski pull requests are the things that we're trying are obviously all of the automated review tools we're putting in place. I think one of the things that we are, and this is kind of goes to one of the things that we're building right now, which is called surfaces, is actually just like understanding and visualizing what is actually happening in a PR in a better way…

Andrew Lisowski

So what constitutes a service? I don't think the listeners might really grock that quite yet.

Tarun Sachdeva

A service in what sense?

Andrew Lisowski

A surface. Like what? Oh, a surface. Yeah. Okay, sorry.

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“If they one-shotted an entire feature, I'm naturally going to be skeptical. Sometimes I'm skeptical and then surprised positively that it did a great job. Sometimes you use a very long session to solve something that seems like it should be a one-shot, and it maybe tells you something about why your code base is hard to reason about. Your agents are having trouble reasoning around your code base in some way.” — Tarun Sachdeva, devtools.fm: Developer Tools, Open Source, Software Development

A host said it's hard to tell a one-shot prompt from a session someone spent an hour on, and a trace shows the difference. Tarun adds that comparing task difficulty with session length can reveal problems in a codebase, not just in how people work.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-02 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 32:14 into the episode
Tarun Sachdeva

…those are the two things that I think are the most important to us and where we've seen the most value. But yeah, it's not quite a perfect marriage of those two tools quite yet. I think there's a lot more points of integration there. I'm not sure if that fully answers your question, but like, yeah.

Justin Bennett

…on this? Like, how many tokens are spent? Like, how long was this actual session? Like, you know, it's a very different story between like, okay, I'm looking at the trace and there was like one prompt at the very top and then you know, like 30 minutes of the agent spinning. Yeah, exactly. Exactly.

Tarun Sachdeva

It's like, yeah, I mean, and sometimes that works. Sometimes that works really, really well. It's tough to make. I haven't seen too many instances of, you know, I guess over time, this is like longer and longer horizon tasks are becoming more and more feasible. It's, you know, it's, it's, I would say it's hard to kind of like, and at least this is not something we do internally. We don't like track how many turns necessarily takes to solve how many types of features and stuff like that. But those metrics in aggregate over time, you probably will want to have some visibility into. And like, you know, we can surface that information, but, you know, knowing how someone arrived at something helps me understand what level of thought and depth they put into that. Andrew Lisowski that appropriate for the task, right? Like, if they one-shotted an entire feature, like I'm naturally going to be skeptical. Sometimes I'm skeptical and then surprised positively that actually it did a great job, right? And there's something to be learned from how that one-shot worked and why it worked. Sometimes you use a very long session to solve something that seems like it should be a one-shot, and it maybe tells you…

Justin Bennett

yeah, I think like a, I don't know, like a, there's a surface here, you know, about like effort, effort spent, or like, I don't know, something.

Tarun Sachdeva

Yeah, yeah, exactly. Exactly. We do have that some of that cost data in tokens in like our token surface, which gives you like turn-by-turn view on like how much every turn cost you. And, you know, reveals a lot of your, you know, things around like model choices and things like that. But yeah…

Andrew Lisowski

yeah, the coding harness I use at work recently added a dollar amount to each like session. And seeing that dollar count go up as I like iterate more and more, I'm like, oh, this will just be like a nice little feature I can fire off. And it was like $400 later. Sheesh.

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“I think engineering and software is the ultimate form of leverage. To only focus on the cost associated with a task is probably not the best form of engineering management, at least at this level of fidelity. I often find myself looking at the cost associated with a task and then just not looking at it, because it's kind of unfair. Sometimes those dollars are spent in learning.” — Tarun Sachdeva, devtools.fm: Developer Tools, Open Source, Software Development

A host had described a coding harness that shows a dollar amount per session, and a feature that ended up costing $400. Tarun argues against managing engineering at per-task cost granularity and explains why Traces doesn't surface token counts by default.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-02 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 36:50 into the episode
Tarun Sachdeva

…you can get to here. That, you know, it's, it's really up to you as an engineering leader or manager or developer to act on that in some way, shape, or form. Like, you know, it's a question on whether or not that's data you should use to change something or it's just interesting to observe. So

Andrew Lisowski

yeah, the coding harness I use at work recently added a dollar amount to each like session. And seeing that dollar count go up as I like iterate more and more, I'm like, oh, this will just be like a nice little feature I can fire off. And it was like $400 later. Sheesh.

Tarun Sachdeva

Yeah. Yeah. You know what? I think this is like a really interesting conversation that you're bringing up, man. Like, I think the idea of attaching dollars to every task at that level of granularity is, you know, for engineering is not like, you know, unless you can also attach the like the dollars that that could potentially generate.

Andrew Lisowski

Yeah.

Tarun Sachdeva

Right. Is not a complete view of the story, right? Like, I think it's, it's really, I think engineering and software is a, it, it's a, it's the ultimate form of leverage. Andrew Lisowski to only focus on the cost associated with a task is, I think, probably not the best form of engineering…

Andrew Lisowski

Yeah, just on time savings alone. Like if I prompted $400 away, but I saved three months of my own life like going down different paths and finding like the dead ends. Like it's kind of hard to quantify that upfront. There's a lot of, yeah.

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“Every single time they use these agents, this is not just a log of your computer's activity. This is decision data. This is your decisions, your team's decisions. And that level of fidelity into those decisions should benefit the teams whose data it is, not just the model providers that are providing the models.” — Tarun Sachdeva, devtools.fm: Developer Tools, Open Source, Software Development

After a host said that seeing a month-long AI conversation would be valuable well beyond coding, Tarun argues that agents are becoming the interface for all knowledge work, and asks who should own and benefit from the record of those sessions.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-02 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 41:26 into the episode
Tarun Sachdeva

…be, there's, there, there's so much nuance to these things. And you kind of have to, you have to like think of a task as not just a singular, like, I did this, I pushed this much code, this is how much it costs. It's like, there's a lot more in the middle there that you, you, you have to consider.

Andrew Lisowski

Yeah, just on time savings alone. Like if I prompted $400 away, but I saved three months of my own life like going down different paths and finding like the dead ends. Like it's kind of hard to quantify that upfront. There's a lot of, yeah.

Tarun Sachdeva

Yeah, exactly. Exactly. This is also why we. We intentionally chose to not surface token counts in the product as a default thing, right? Like tokens are not something we are overly fixated on as a product. We are very much of like a tool to help developers collaborate with each other in this new way. And I just find that like token counts are useful in some sense, but like it's so different in every harness and every model. It's just like, how useful is that number really? Is like attaching the number of bytes your code base like consumes. It's like, okay, cool, useful in some contexts, but for the most part, it's just not useful for workflow.

Andrew Lisowski

Yeah. Recently, Dan Abramov, a famous React guy, took a detour for a month to go into mathematics and prove like theorems and stuff. And talking about this makes me think that that type of work really fits to this model too, because he spent a month prompting LLMs and probing the depths of them and…

Tarun Sachdeva

Yeah, exactly. And those conversations are some of the most interesting. You know, we have some really interesting ones on our that are shared publicly as well. And I think that, you know, your point here is exactly very astute is that agents are the primary interface to really all knowledge work.…

Justin Bennett

There's something that I would, I'd love to sort of like chat through. As you were talking about like thinking of like these certain features as like a call center or something, there's like, you know, classic examples of like engineering orgs getting treated as call centers and that like really…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“When we started working through agents, the one thing that we kind of lost visibility into was what happened between you and the code. Before, there was no distance. It was just you and your editor, and you would put it on GitHub and collaborate there. There was this whole in-between thing, a back and forth of you and the agent, that was just no longer being captured.” — Tarun Sachdeva, devtools.fm: Developer Tools, Open Source, Software Development

Asked what need led him to build Traces, Tarun describes the shift to agents as the main way code gets written. His team could still review pull requests but could no longer see how anyone was prompting the agent.

devtools.fm: Developer Tools, Open Source, Software Development · 2026-11-02 Permalink → Listen →
devtools.fm: Developer Tools, Open Source, Software Development Around 10:22 into the episode
Tarun Sachdeva

…It's been really interesting to like apply the learnings of one type of community in a completely different domain and also just see what is common across all of them. So, yeah, it's I've been very fortunate to have worked on some like really, really interesting problems in those kinds of spaces.

Justin Bennett

…previously that like Traces sort of spun out of the work that you were doing. Yeah. Previously, it's like you realized that y'all didn't have the tools you needed, you needed something else. What did you need? And what did you sort of like set out to accomplish when you were building Traces?

Tarun Sachdeva

Yeah, so when we, I mean, you know, this, I think this is like the Opus 4.6 moment about a year ago today, now it's like October last year, when we really started to see that we were moving beyond the primary way of using AI and coding was for me that switch was moving beyond autocomplete into the agent entirely, right? Like was really around this time. And when that happened, when the agent became the sort of primary interface to how code was being generated, you know, we noticed that one of the key things that we used to do as a team, and we were a small team then, we're still a small team now, but we're a little bit bigger, but still, still a small team. What we used to do was we used to code, we used to write code and then collaborate on GitHub and talk about our things around pull requests, right? Like that's kind of how our workflow worked, just like everyone else, right? And we would provide input into each other's work in that way. And when we started working through agents, the one thing that we kind of lost visibility into was this like what happened between you and the code, right? Where before there was no distance, it was just you and…

Justin Bennett

Nice. That's awesome. So for. For our audience who may not have experienced your product yet, would you tell people a little bit more about just what is the experience that you're providing?

Tarun Sachdeva

Yeah, yeah, that's great. So, we intentionally shaped traces to feel familiar, right? We want it to feel familiar. And the product that we often compare to is GitHub, right? So, when you join traces.com, you get a personal traces account. And you get traces.com/slash your name. So, you can grab…

Andrew Lisowski

Do you have users that like are default to public? Because, like, when I use GitHub, it's like I have one decision at one point in time. I want this repo to be public, and then all pushes therefore after are public. Like, I feel like it might be a little cumbersome to be like, okay, every single…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

Collections They Appear In