Mar 14, 2026 · 34m · latent-space
⚡️Monty: the ultrafast Python interpreter by Agents for Agents — Samuel Colvin, Pydantic
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Pydantic creator Samuel Colvin joins Swyx to unveil Monty, a high-performance in-process Python interpreter built in Rust for AI agent execution. The discussion covers runtime sandboxing trade-offs, practical techniques for 100x developer productivity with AI coding assistants, and full-stack agent observability using Pydantic Logfire.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 17.5% of the talking time here. How this is scored →
The hosts as informed peer 5.0 Guest teaching 4.0 Guest disagreement 1.4 The hosts pushing back 2.8
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 7:00 Dismissing Pyodide for server-side code execution
Samuel bluntly labels relying on Pyodide outside the browser as a classic mistake and outlines its fatal sandboxing and memory management flaws.
Hardest push from the hosts ▶ 30:24 Refusing TOML as an agent specification standardSwyx directly disputes the serializable TOML format, arguing that markup configurations inevitably degrade into poorly implemented general-purpose languages.
Biggest teaching moment ▶ 7:15 Technical deep dive into Deno and Wasm isolation risksSamuel educates the audience and Swyx on why WebAssembly runtimes inside Deno fail to guard against out-of-memory crashes and cross-invocation server tainting.
The host holds their own ▶ 19:52 Correcting assumptions about Google's internal code trainingSwyx demonstrates domain expertise by countering the guest's theory, explaining internal Google toolchains like Jet Ski and why direct training on Borg code would yield nonsense.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome Back and Pydantic Logfire AI Observability | 4 | 3 | 1 | 2 | Swyx opens the interview smoothly and probes the specific meaning of an AI-native observability platform. Samuel explains Logfire's architectural design choice to expose raw SQL querying over OpenTelemetry traces. | |
| The Origin of Monty and the Need for Code Mode | 4 | 4 | 1 | 1 | Samuel relates how Anthropic researchers inspired Monty and contrasts code-mode execution directly with heavyweight VM sandboxes like Daytona and Modal. Swyx validates the dynamic between builders and marketers. | |
| Evaluating Runtime Trade-offs and the Limitations of Pyodide | 5 | 6 | 2 | 4 | Samuel details why Pyodide and Deno sandboxing failed security and latency requirements for backend execution. Swyx pushes on the paradoxical nature of a Pydantic runtime unable to run Pydantic. | |
| Accelerating Development: The Four Rules for 100x AI Speed | 5 | 5 | 1 | 2 | Samuel outlines his four rules for achieving 100x acceleration when having LLMs write reimplementations like Python built-ins. Swyx inquires about the internal GitHub importance scoring badge and automated code reviews. | |
| Comparing AI Coding Agents: Claude Code, Codex, and Gemini CLI | 7 | 4 | 2 | 5 | Samuel compares Claude Code, Codex, and Gemini CLI, speculating that Gemini excels because it trained on Google's internal monorepo. Swyx pushes back using firsthand knowledge of Google's internal Jet Ski and Borg infrastructure. | |
| Live Demo: Secure Agentic Web Scraping with Monty and Logfire | 3 | 5 | 1 | 1 | Samuel conducts a walkthrough demonstration of an autonomous web scraper built using Playwright, BeautifulSoup, and Monty with Logfire tracing. Swyx follows along as an appreciative audience. | |
| Standardizing Agent Architecture with Serializable Agents | 7 | 3 | 2 | 6 | Samuel proposes TOML-based serializable agents as a declarative standard. Swyx challenges this framing, pointing out that configuration DSLs inevitably devolve into half-baked programming languages. | |
| Industry Conferences, Monetization, and Concluding Thoughts | 5 | 2 | 1 | 1 | The conversation winds down with promotions for upcoming London and Pydantic conferences. Swyx draws parallels to runtime evolutions like Bun and Deno while discussing monetization paths. |
Statements from this episode (16)
Colvin wrote 30,000 lines of Rust for Monty over Christmas using AI
“This is a crazy world where you can sit down over Christmas and write 30,000 lines of rust that, that is actually powerful and useful.”
Colvin: Type safety is critical for AI tool execution
“I think type safety is important for humans, but it's critical for AIs.”
Colvin: Monty executes Python in single-digit microseconds
“I mean, in a hot loop, we can run, we can go from code to execution result in under a microsecond in like 800 nanoseconds. In reality, it's like single digit microseconds to run code or single digit microseconds to run the next step of a REPL or single digit m…”
Colvin: Deno cannot control memory usage and prevents OOM protection
“Even if you don't allow that Dino does not have any way of controlling memory. So even if someone can't run arbitrary code, they can oom your machine as often as they like.”
Colvin: Monty will never support CPython ABI packages like NumPy
“There's no support for third party libraries just to be installed. There never will be directly as, and you'll never, we'll never be able to speak the C Python ABI and like install Pydantic or install NumPy or something.”
Colvin: LLMs are 100x faster under four specific engineering conditions
“My take is that there are four, four things where if you can cover all four of these things, LLMs are not like three X faster or five X faster. They're like a hundred X faster.”
Pydantic built a VIP issue scorer after closing an OpenAI founder's ticket
“Basically this started off because one of the OpenAI co-founders created an issue on Pydantic. And we just closed it and said it was wrong. And so we have this that, like, injects itself and tries to summarize someone and it gives them a, like, brutal score of…”
Colvin: Pydantic team uses Devin after finding it best off-the-shelf coding agent
“We, so in Pylandsk AI, the guys are using Devon quite a lot. And actually we tried quite a few, nothing else worked, but Devon, well, not nothing else worked, but like having gone through a few different options, Devon seemed to be the best of the off the shel…”
Colvin uses Gemini CLI to review branches and Claude Code to implement
“Gemini doesn't, is not allowed to by default run, like edit any files and it will just go off and like review a particular branch and write me out a report. And then I basically just point called code at that report and say, implement the following things.”
Walmart sees enormous internal adoption of Pydantic AI coding agent CodePuppy
“CodePuppy is a coding agent built by an amazing guy called Michael Faffersberger at Walmart built with Pydantic AI. And it's got like, hasn't got very many stars. Hasn't got lots of like hype around it, but I don't know how much I'm allowed to say publicly, bu…”
Google developers internally use a coding tool called Jet Ski
“They use an internal thing called jet ski. And I mean, the simple reason is they have internal versions of everything else. And so if they actually train the, they release the internal version to the external world, it would just make no sense. Like it would j…”
Colvin: Monty embeds the TY type checker before running code
“Monty has the TY type checker built into it. So before it will go and run any code, It's running type checking with these type stubs.”
Colvin: Untrusted users prompting cloud AI is equivalent to letting them write code
“If you're running this kind of thing in the cloud and you, and you're gonna have ultimately untrusted people prompting the model, that is effectively the same as letting an untrusted person write the code.”
Colvin: There is massive opportunity in an agent optimization state layer
“I think that there is an enormous opportunity for effectively a new layer of state within applications. You could think of it as memory, but it's often a lot more than memory. It's the like current state of agent optimization, some of which will be code, some …”
Colvin: Pydantic AI introduces declarative TOML serializable agents
“One of the things we're doing now in Pydantic AI is where we're about to introduce, I think there's a PR out for this. So I think this is public serializable agents. So basically you can define an agent entirely in a TOML file, everything from the model to the…”