I believe several!
Replying to @dgerlanc
30 out of 100 is huge. asking about their problem with no pitch is what makes it work, people love explaining their own problem. did any of those 30 end up as the first paying customers?
58
“Yeah, it was literally like, we have nothing. We have no product. I saw you have this problem. Can we help you solve it…And I emailed about 100 founders and about 30 actually emailed me back, which is crazy high.”
On the latest episode of the Agents + Engineers podcast, @suchintan and I discuss what it took to start from a blank slate to building a full agentic browser automation platform.
Suchintan is a co-founder of Skyvern, which uses AI agents and computer vision to automate browser workflows. Previously, he built machine learning platforms at Faire and Gopuff.
We also discuss:
- How early customer pilots shaped Skyvern, and why the team decided to open-source it
- Why Skyvern rewrote its browser agent as models improved, and how it reduced the cost of running automations
- What browser agent benchmarks can tell us, and why strong scores don’t always translate into useful products
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/2rz…
"It's all comes down to hardware. And I think that's the fun thing about the AI era...Like there's no way around the fact that even though we're writing all this software at the end of the day, we're constrained by hardware."
Erica Hughberg is a software engineering leader, speaker, and community advocate at Tetrate. She’s also a maintainer of Agent Router, formerly Envoy AI Gateway.
In this episode, we discuss:
- What Agent Router does and how AI workloads challenge gateways and proxies
- Why coding agents can struggle with undocumented decisions and the context of an existing system
- How coding agents can help people apply their expertise in unfamiliar programming languages
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/6ZE…
2
134
"Tt's one of those things that you realize it's really just learning from your existing code. It doesn't really know what's good. It thought you've seen this code pattern before, this is the right way...and it really needs a human to know that's not how you're supposed to be doing things."
AI agents can write code quickly, but they can also repeat bad patterns and produce contributions that maintainers never asked for. Python core developer Mariatta Wijaya uses agents in her own work while seeing the challenges their use creates for open source projects.
Mariatta is an open source maintainer and contributor with over 20 years of software engineering experience. As a Python core developer, she focuses on improving the core workflow and documentation. She’s also an advisor and administrator for Global PyLadies, a co-founder of PyCascades, and a former chair of PyCon US, PyCascades, and PyLadiesCon.
In this episode, we discuss:
- How AI-assisted contributions affect CPython maintainers, and what new contributors learn by participating in open source.
- How Mariatta uses plans, documentation, and agent skills to guide coding agents and correct recurring mistakes.
- Why she builds personal apps with agents, from a scheduling tool to a workout app.
2
84
“And I think like humans, we should spend a lot of our time on the beginning part. So the one slice of bread, which is thinking about the idea, the design, the requirements, what does success look like? The middle is what is totally automatable. The AI agents go build it. And then the other half of the sandwich, the bread on the other side is the polish part.”
On the latest episode of the Agents + Engineers podcast, @trevin and I discuss where we should spend our time as builders when agents can write all the code.
Trevin is a product leader and builder with more than 25 years of experience in product and design, including leadership roles at Axon, Sketchy, Nike, and Microsoft. He’s also a core maintainer of Compound Engineering, a plugin for Claude Code, Codex, and other agent harnesses.
We also discuss:
* How Trevin, starting with a product’s goals, uses agents to explore designs before generating prototypes.
* What it takes to build and test skills that work across different models and coding agents.
* Why sharing documents and context between agents is still difficult, and how Trevin’s open source project Energon addresses it.
Listen directly to the episode and be sure to rate and subscribe to the show!
apple: podcasts.apple.com/us/podcas…
spotify: open.spotify.com/episode/2eD…
5
6
164
Dan Gerlanc retweeted
Vycari treats design docs as live inputs: nightly checks compare them with code and suggest updates, while detailed PR follow-ups become issues for the implementation pipeline.
agentsandengineers.com/episo…
"I think the big thing that we're starting to realize is that the monorepo isn't serving us very well. And so we've started to really aggressively split microservices out of our monorepo into their own repos with their own CIs."
On the latest episode of the Agents + Engineers podcast, @allen_hutchison and I discuss what it’s like going from leading large organizations at Google to building a startup with a co-founder and more than a few agents. We walk through Allen’s agent setup, including the engineering decisions that still need human attention.
Allen is the co-founder and CTO of Vycari, where he builds AI agents for everyday tasks. He previously spent 21 years at Google, most recently as a Distinguished Engineer at Google DeepMind working on AI agents and developer tools. He was part of the founding team behind Gemini CLI and built the Antigravity SDK.
We also discuss:
- How Vycari uses design docs, nightly audits, and agents that pick up follow-up issues to keep development moving
- Why Allen is building agents into familiar interfaces like calendars and to-do lists
- How he chooses models for different tasks, including using more capable models to write skills for less expensive ones
Listen directly to the episode and be sure to rate and subscribe to the show!
“I know it's far slower to do it that way, but if I want people to be able to look at the code I've written and use it as an example, I want it to be, you know, everything's there for a reason. I can explain it all.”
On the latest episode of the Agents + Engineers podcast, Ben Morss and I discuss both how AI has vastly accelerated our ability to write code and solve problems. However, for learning, understanding, and teaching, there is still no substitute for writing the code yourself.
Ben is a Developer Advocate at Mistral. He previously spent 9.5 years at Google, where he was a Product Manager for advanced APIs in Chrome and a Developer Advocate for the web. He earned a bachelor's degree in Computer Science and Music from Harvard University and a doctorate in Music Composition from the University of California at Davis. He continues to make music and write about music theory.
We also discuss:
- Why AI still can’t do music theory
- The challenges of out-of-distribution training data, even for modern AI
- The human/AI balance in problem-solving with people
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/1Rt…
YouTube: youtu.be/WCNox1rAytA
a+e website: agentsandengineers.com/episo…
“How do we get to the point where I can say the business analyst, Margaret in accounting, Tony in sales, Jana in marketing, why can't they deploy 10 times a day?”
On the latest episode of the Agents + Engineers podcast, Cory O’Daniel and I discuss how the role of software engineer changes when people who aren’t traditional software engineers, or, as we’ll call them, citizen developers, can, with an agent and enough patience, create useful software.
Cory O'Daniel is the CEO and co-founder of Massdriver, an internal developer platform and platform orchestrator that helps teams turn Terraform, OpenTofu, and Helm into governed self-service infrastructure. A software engineer and cloud-operations veteran, he has spent more than 20 years building teams and startups and working with cloud infrastructure.
We also discuss:
- Thinking about identities and permissions for agents
- Testing cloud infra rollouts in an agentic loop, and
- How we’re all citizen developers in one way or another
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/3NN…
1
94
Dan Gerlanc retweeted
Revisting this. Big thanks @dgerlanc for having me on the podcast, good dialog about all kinds of things agentic.
As the clip shows, I want to dive into the claim that AI will make Python irrelevant, that Python is a poor implementation choice once the agent does the coding for you. Some holes in that story.
“We're going to have humans in the loop, and humans are going to need to read code and sign off on it before shipping it. And readability will still matter, and Python remains the hands-on winner for that.”
On the latest episode of the Agents + Engineers podcast, Paul Everitt, Developer Advocate at JetBrains and I discuss whether humans will still be reading the code the agents create.
I’ve seen a lot written about this topic lately. Yes, you need to verify code works regardless of how it’s created. But are you willing to submit an agentically developed PR to the Linux kernel without reading the code?
We also discuss:
- How to create handoff packets between different model sizes for efficient delegation
- The state of the art for running local models
- What agentic engineering means for Python development
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/6ZT…
a+e Website: agentsandengineers.com/episo…
How does Gas City, a popular open source framework for building your own software factory, manage its own development? In the latest episode of the Agents and Engineers podcast, we’ll talk with one of the maintainers of Gas City on how she, and the project, approach agentic development. We’ll also discuss ways to customize your own Gas City deployment.
@sgjarmak is an AI Engineer at Omni working on multi-agent orchestration and code intelligence. She is also a research affiliate for NASA's Science Explorer (SciX) and a maintainer of Gas City, an open-source orchestration builder for multi-agent coding workflows.
In this episode, we discuss:
- How Stephanie went from doing a PhD in physics and working for NASA to joining a tech company
- How Stephanie uses AI to create fun and educational games for her kids
- The agentic pre-mortem as a strategy
Listen directly to the episode and be sure to rate and subscribe to the show!
apple: podcasts.apple.com/us/podcas…
spotify: open.spotify.com/episode/3ws…
a+e website: agentsandengineers.com/https…
Dan Gerlanc retweeted
I recorded a podcast with @dgerlanc last week where I talk about my experience with agentic development, software factories, OSS maintaining, and more! agentsandengineers.com/episo… (pictured: a Cairn, the name my Mayor chose)
Not sure if they’re still available. Even years ago the logits were truncated so you couldn’t normalize.
You could test with frontier open source models. My guess will be that they don’t do much better!
Replying to @dgerlanc
Yup. I want to know which domains are good ones for Jev. So far it’s good at type 1 thinking style questions. But… is it really better than what you would get from the current frontier models if we had access to their logprobs? And maybe we already do. I haven’t checked in a while.
1
151
Your reasoning is correct (60/40)! Just I wouldn’t expect Lev to be able to produce a meaningful probability estimate from the censored logits. It’s some kind of joint probability of the tokens, which doesn’t necessarily map to the symbolic probability of the correct answer, if that makes sense.
Replying to @dgerlanc
Really? 🤔 If told you that I have an unfair coin that comes up heads 60% of the time and asked you to ascribe probabilities to `heads` vs `tails` for the next flip, then what you say?
Perhaps I misunderstand what you mean by "calibrated" here.
1
208
“If you're not throwing a bunch of things away every week and starting over or adding new stuff or simplifying ever like every week, then you're probably not iterating fast enough”
On the latest episode of the Agents + Engineers podcast, @andrewzigler and I discuss the importance of continue experimentation to improving agentic workflows.
Andrew is a GTM Engineer at LinearB and the host of Dev Interrupted. He has written extensively about agentic engineering, including his technique of Mise en Place.
We also discuss:
- Why individual gains from AI coding often fail to become organizational gains
- Andrew’s technique of Mise en Place + Beads for agentic dev
- The value of a humanities education in an agentic world
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/2Mk…
a+e website: agentsandengineers.com/episo…
Dan Gerlanc retweeted
The Industrial Revolution gave us machines that could mangle our bodies. The AI revolution is giving us machines that can mangle our minds.
After a day of Claude Code, dopamine loops, and "agentic gooning", I wonder whether dark software factories might be a form of workplace safety.
The thought experiment is simple: agents write and ship the code, and humans aren’t allowed to look at it. You have to figure out what evidence you need to know the software works.
That forces some uncomfortable questions. What are your tests actually checking? How do you verify behavior? What would give you enough confidence to put something into production without reading every line?
VPs of Engineering have managed versions of this for years. They don’t read everything their teams ship. They rely on people, tests, and systems of verification, while taking responsibility for what gets built.
As agents produce more code than we could possibly read, we’re going to need much better versions of those systems.
But I also wonder what happens to us when we stay inside the loop all day: prompting, checking, getting another little hit when something works, pushing further, and eventually crashing.
And we’re doing all of this while marinating in the daily Dario/Altman-induced psyops.
We learned to put distance between people and dangerous industrial machinery. Could we design software production with some distance between our minds and the machinery too?
I talked about this with Dan Gerlanc (@dgerlanc) and Doug Turnbull (@softwaredoug) on Agents and Engineers, including why Terence Tao’s questions about machine-generated proofs belong in this conversation.
What would you need to see to trust software you weren’t allowed to read?
Full episode:
agentsandengineers.com/episo…
Think this is a good demonstration that you CAN just use Python. When you have millions of RPS, then optimize. Most services will never get to that scale.
Habitat is OpenAI's online storage platform that powers everything from ChatGPT to Codex.
It has grown over 10x year over year. Before the Rust rewrite, its Python service handled more than 20 million requests per second at peak.
1
109
“The job of software engineering is going to become much, much, much more interesting. You're basically going to be left with the interesting problem, the hard stuff, the stuff that requires taste.”
On the latest episode of the Agents + Engineers podcast, Julien Verlaguet and I discuss where software engineering is headed with the release of increasingly capable models.
Julien Verlaguet is the founder and CEO of SkipLabs, the company building Skipper, a closed-loop coding agent. He created Skip, a reactive programming language, and led the design of Hack, the language Meta developed to run its codebase at scale.
We also discuss:
- What do programming language designers do when most code is being written by agents?
- How Julien refocused SkipLabs towards AI
- The use of formal methods with AI
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/4j7…
a+e website: agentsandengineers.com/episo…
“A premise we had like, I don't know, 15 years ago, we're like, oh yeah, data science is going to be like this amazing thing that's like going to deliver all this business value. I don't think we have really delivered. I think for a large part we've failed, right?”
On the latest episode of the Agents + Engineers podcast, @twiecki and I discuss how agentic AI has finally allowed for data science to deliver on its original promises of business value.
This is not the say that data science failed universally, more that the amount of engineering resources needed to deliver data science in production was often beyond the grasp of all but the largest organizations.
Thomas Wiecki is a co-author of PyMC and the founder of @pymc_labs. He holds a PhD in Computational Cognitive Neuroscience from Brown University and previously served as VP of Data Science and Head of Research at Quantopian Inc.
In October, Thomas is also teaching a course, Master Agentic Data Science, where he’s teaching hands-on some of the things we discuss in the episode.
Hugo Bowne-Anderson, who was on the most recent episode of the podcast, is also one of the instructors.
You can find a link to sign up for the course here: pymc-labs.com/courses/agenti…
We also discuss:
- How Thomas’s work in grad school building models of the brain led him to Bayesian statistics
- Effectively embedding agents in tools like Slack or Discord to democratize access to company information
- What parts of doing data science still require talking with humans
Listen directly to the episode and be sure to rate and subscribe to the show!
Apple: podcasts.apple.com/us/podcas…
Spotify: open.spotify.com/episode/7on…
a+e website: agentsandengineers.com/episo…










