2026-06-29 · I'm open-sourcing Bottega, our internal coding agent orchestration tool

Transcript

Transcript. Paraphrased; sources in notes.md.

HostHello everyone, and welcome to the show. It is June 29th, 2026, and today we have a packed slate. We are digging into the shifting landscape of agentic workflows, some new tooling to help us optimize our local hardware for LLMs, and finally, a deep dive into running models natively within OCaml. Samantha, it is great to have you back. How are you feeling about the state of our build environment today?

GuestDaniel, I am feeling dangerous. I spent half the morning looking at how people are trying to automate their entire PR process, and honestly, the sheer amount of configuration some teams are putting into these workflows is wild. It makes me want to rip out half my stack just to see if it still runs.

HostThat actually leads us right into our first story. There is a project called Bottega that has just gone open-source, and the author is claiming that after shipping over 1,000 user stories with their internal agent orchestration tool, they have moved to a state where 100 percent of their production code is written by agents. The core of their argument is that we need to stop treating agents like magic wands and start treating them like junior developers who need an ironclad planning stage and a rigid review loop.

GuestI love the honesty here. Everyone is chasing that holy grail of just typing a prompt and getting a feature, but the reality is that the mess usually happens in the back-and-forth iteration. If you do not have a formal plan or clear acceptance criteria before you even let the model touch a file, you are just signing up for a headache. I have been advocating for this kind of rigorous gating for months. It is not about killing creativity, it is about keeping the context window from turning into a circular argument.

HostIs it too rigid, though? The author notes that they built this specifically to address the fatigue of dealing with giant, unreviewable pull requests.

GuestIt feels like the right kind of friction. If the alternative is babysitting an agent for three hours to fix a single bug because it wandered off into a hallucinated dependency, I will take the planning phase every time.

HostMoving from process to hardware, our second topic is a tool called llmfit. This is a terminal utility that benchmarks your local machine, checking your RAM, CPU, and GPU stats to recommend exactly which LLM models are going to perform best on your specific hardware.

GuestThis is pure gold. How many times have you pulled down a new model, spun it up, and then realized your machine is sweating like a marathon runner on mile twenty? llmfit, for the uninitiated, is a command-line tool that profiles your hardware and tries to match you with models that actually fit in your memory ceiling. I used something similar last week to profile an internal dashboard I was building, just to make sure I wasn't blowing through my VRAM every time I hit a refresh. It is practical, it is fast, and it stops me from trying to run 70-billion-parameter models on a machine that clearly cannot handle the load.

HostIt is a classic problem. We keep getting these massive new models, and unless you are running a rack in your basement, you really have no idea if you are going to get five tokens a second or thirty until it is already crashing your session.

GuestExactly. And the TUI interface they built is clean. It turns that opaque mystery of model hardware requirements into something you can glance at while you are still deciding which weight to pull down.

HostFinally, let's talk about the OCaml ecosystem. There is a new note from the community about integrating LLMs directly as functions within OCaml code, bridging the language with Antirez's Dwarfstar inference engine. Antirez, for those who don't know, is Salvatore Sanfilippo, the original creator of Redis, and Dwarfstar is his lightweight, native inference engine designed for portability. The author’s claim is that by treating the LLM as a library rather than calling out to a bloated command-line interface, you get a much tighter, more predictable agentic loop.

GuestThis is the vibecoding evolution I have been waiting for. We are moving away from these sprawling, external CLI tools that hide all their logic in a black box, and toward native, type-safe function calls. When you pull the model into the language runtime itself, you gain so much control over the execution flow. It is elegant.

HostDo you think this is going to stay a niche for functional programming enthusiasts, or is this the way we should be handling agentic integration everywhere?

GuestI think we are heading for a split. You are going to have your high-level, "just make it work" tools for the quick wins, and then this approach for the systems where you actually need to understand the agent's state at every single step. It is the difference between writing a script and building a real, hardened piece of software. I think a lot of people are going to realize they have been doing it the hard way once they see how clean that OCaml implementation is.

HostIt is a great point to end on. It feels like we are finally starting to treat these agents with the same respect we give to any other complex dependency.

GuestAbsolutely. I might spend my Sunday trying to map out a similar function-first approach in my own projects. Maybe then I will finally get those PR review times down to something that doesn't feel like a full-time job.

HostThat sounds like a solid plan. Thanks for walking us through all of this, Samantha. All of today's stories and the broader discussions can be found over on Lobsters. Thank you all for listening, and we will see you back here tomorrow.