May 7, 2025 · 1h 16m · latent-space
Claude Code: Anthropic's CLI Agent
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
In this Latent Space episode, Anthropic's Kat Wu and Boris Cherny discuss Claude Code, exploring its minimalist terminal architecture, internal dogfooding genesis, economic return on engineering productivity, and safety paradigms for agentic coding.
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 33.7% of the talking time here. How this is scored →
The hosts as informed peer 5.9 Guest teaching 5.0 Guest disagreement 1.3 The hosts pushing back 2.0
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 45:27 Challenging external memory architectures
Boris directly counters conventional AI agent memory architectures, arguing that everything will eventually be subsumed directly into the foundation model.
Hardest push from the hosts ▶ 45:44 Pushing back on infinite context vs agent controllabilitySwyx rejects the idea of delegating all context to the model, insisting developers need explicit external constraints and auditability over what an agent knows.
Biggest teaching moment ▶ 48:00 Deconstructing vector RAG vs agentic searchBoris explains how agentic glob and grep code search decisively beat vector databases by eliminating index synchronization latency and third-party security vulnerabilities.
The host holds their own ▶ 23:20 Dissecting CLI architecture and React Ink decompilationSwyx demonstrates deep domain knowledge from his background maintaining the Netlify CLI, citing Commander JS, Vadim Demedes' React Ink, and Bun compilation.
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 and Introductions to the Latent Space Studio | 4 | 3 | 1 | 1 | Hosts warmly welcome Boris and Kat, noting shared history at Dagster and coffee shop recognition. Boris explains how Claude Code started organically inside Anthropic as a scrappy experimental research tool. | |
| Anthropic's Product Philosophy: Simple Solutions and Forward-Looking Models | 5 | 4 | 1 | 2 | Swyx probes on product management velocity and planning for model improvements 3 months out. Kat and Boris outline Anthropic's philosophy of doing the simple thing first and building for future model autonomy. | |
| Scaffolding Architecture, Context Compaction, and CLAUDE.md Configuration | 6 | 5 | 1 | 2 | Alessio queries prompt optimization and scaffolding thickness compared to Cursor. Boris explains that context compaction was implemented simply by prompting Claude to summarize past turns rather than over-engineering memory structures. | |
| Aider Lineage, AGI-Pilled Moments, and Composable Unix Workflows | 6 | 5 | 1 | 2 | Hosts ask about Aider comparisons and tool positioning across IDEs and terminal agents. Boris shares his initial AGI-pilled moment with internal tool Clyde and frames Claude Code as a composable Unix utility. | |
| Token Economics, Engineering ROI, and Model Selection Strategies | 5 | 6 | 3 | 3 | Alessio contrasts Cursor's flat fee with pay-per-token API costs. Kat and Boris reframe the discussion away from raw token cost toward engineering ROI given typical daily user expenses around six dollars. | |
| Major Feature Ships and Dogfooding Claude Code's Implementation | 5 | 4 | 1 | 1 | Kat recaps recent feature additions including WebFetch, auto-compact, auto-accept, and vim mode. Boris reveals that around 80 to 90 percent of Claude Code itself is generated by Claude with human code review. | |
| Custom Slash Commands, MCP Integration, and Semantic CI/CD Linting | 6 | 5 | 1 | 2 | Alessio explores the conceptual boundary between local slash commands and MCP integrations. Boris provides a concrete case study of using Claude Code in GitHub Actions for semantic linting that static analyzers miss. | |
| Technical Underpinnings: React Ink, Terminal ANSI Standards, and Bun | 8 | 3 | 1 | 2 | Swyx brings deep domain expertise from maintaining the Netlify CLI and references React Ink and Bun binary packaging. Boris details the friction of 1970s ANSI escape codes and cross-terminal rendering quirks. | |
| Safety Guardrails, Permission Boundaries, and Autonomous Safety Levels | 7 | 5 | 2 | 4 | Alessio asks why file writing is considered unsafe if version control exists. Boris clarifies risks around prompt injections and early failure detection, while Swyx cites the METR benchmark on human intervention intervals. | |
| Headless Non-Interactive Mode and Automated Test Suite Remediation | 6 | 6 | 1 | 2 | Kat and Boris explain non-interactive mode with the dash-p flag and tool allow-listing for headless workflows like PR changelogs and flaky test remediation. Kat advises starting with single-test verification before scaling. | |
| Engineering Leadership, Quality Standards, and Automated Unit Testing | 6 | 5 | 1 | 2 | Swyx asks how engineering leadership manages code review standards when developers generate huge PR volumes. Kat emphasizes developer accountability while Boris notes he no longer writes unit tests manually. | |
| Productivity Measurement, Slack Triage Bots, and Agile Prototyping | 6 | 5 | 1 | 2 | The conversation turns to productivity metrics like cycle time and PR throughput. Boris highlights an internal bot that parsed Slack bug reports and automatically submitted functional PR fixes. | |
| Internal Tooling Creation and Multimodal Terminal-to-Browser Prototyping | 5 | 5 | 1 | 1 | Kat discusses the boom in internal Streamlit dashboards enabled by zero-to-one generation. Boris describes feeding UI design screenshots into Claude Code and having Puppeteer autonomously iterate on frontend alignment. | |
| Agent Memory Architectures, Context Windows, and Interpretability | 7 | 6 | 3 | 4 | Swyx and Boris debate external memory structures versus pure model context. Swyx pushes back against blind model trust, arguing for auditable limited memory, while Boris points to mechanistic interpretability research. | |
| Why Agentic Code Search Replaced Traditional Vector RAG | 7 | 7 | 2 | 2 | Boris explains why Anthropic abandoned vector RAG in favor of agentic glob and grep search, citing quality, sync drift, and security. Alessio drills into the exact distinction between the Think Tool and extended chain-of-thought. | |
| Sandboxing Environments and Parallel Sub-Agent Exploration | 6 | 6 | 1 | 2 | Boris explains using sub-agents to explore parallel architecture paths simultaneously. Kat analyzes how Sonnet 3.7's relentless persistence sometimes causes it to take instructions too literally compared to earlier models. | |
| Cross-Session Continuity, Git History, and Worktree Execution | 6 | 5 | 1 | 1 | Hosts ask about preserving context across terminal sessions. Boris and Kat explain using git diff tracking and git worktrees for isolated parallel sub-agent workspaces. | |
| Pre-Commit Hook Philosophy and Rapid Zero-Shot Markdown Parser Creation | 6 | 5 | 1 | 1 | Swyx brings up team debates around pre-commit hooks. Boris shares a story of building a custom, terminal-optimized Markdown parser via Claude Code at 10 PM on the eve of release when off-the-shelf libraries fell short. | |
| Enterprise Commercialization, Pricing Models, and Productivity Multipliers | 5 | 5 | 1 | 2 | Alessio and Swyx ask about enterprise monetization plans and productivity metrics. Boris estimates a 2x personal speedup and notes top engineers seeing up to 10x gains on specific workloads. | |
| Empowering Non-Technical Users and the Future of Prompt Engineering | 6 | 5 | 2 | 3 | Boris shares examples of non-technical designers and finance analysts piping CSV data into Claude Code. Alessio and Boris discuss whether prompt engineering will remain essential or get erased by model reasoning. | |
| Open Source Considerations, Continuous Refactoring, and Terminal UX Design | 6 | 5 | 2 | 2 | Swyx asks why Claude Code is closed source and suggests source-available licenses. Boris discusses maintenance overhead and reveals the codebase is rewritten as a Ship of Theseus every few weeks. | |
| Anthropic's Developer Focus, Foundation Strengths, and Team Recruiting | 5 | 4 | 0 | 1 | Swyx asks why Anthropic became the developer favorite for coding models. Kat highlights software engineering as the ideal high-ROI domain for LLM agency, and Boris closes with recruiting invites. |
Statements from this episode (44)
Boris Cherny: Claude Code was prototyped using Anthropic's public API
“When I joined Anthropic, I was experimenting with different ways to use the model kind of in different places. And the way I was doing that was through the public API, the same API that everyone else has access to.”
Cherny: Internal DAU for Claude Code went vertical across Anthropic
“Then we gave all the engineers and researchers at Anthropic access and pretty soon everyone was using it every day. And I remember we had this DAU chart for internal users and I was just watching it and it was vertical like for days”
Anthropic avoided building an IDE to target future AI model scaling
“If we want some product that has like very broad a product market fit today, we would build, you know, a cursor or a Windsurf or something like this. Like these are awesome products that so many people use every day. I use them. That's not the product that we …”
Anthropic compacts Claude Code context by having Claude summarize older messages
“We tried a bunch of different options for compacting, you know, like rewriting old tool calls and truncating old messages and not new messages. And then the end, we actually just did the simplest thing, which is ask Claude to summarize the, you know, the previ…”
Cherny: Highly capable models make simple scaffolding sufficient
“And it's funny with, when the model is so good, the simple thing usually works. You don't have to over-engineer it.”
Anthropic implements Claude Code memory as a plain CLAUDE.md auto-read file
“And WhatMD, it's another example of this idea of, you know, do the simple thing first. We had all these crazy ideas about, like, memory architectures, and, you know, there's so much literature about this. There's so many different external products about this,…”
Cherny: Aider inspired internal Anthropic tool Clyde, which inspired Claude Code
“It was, ah, Adr inspired Clyde, which inspired Cloud Code.”
Cherny: Some Anthropic engineers rack up thousands daily running Claude Code automations
“And there's some people at Anthropic that have been racking up like thousands of dollars a day with this kind of automation.”
Cherny: Claude Code is built as a composable Unix utility like grep
“We think of it as like a Unix utility. Right. So it's like the same way that you would compose, you know, grep or cat or oh, cat. Or something like this. The same way you can compose code into workflows.”
Anthropic observes Claude Code API costs averaging roughly $6 daily per user
“Currently we're seeing costs around, like, six dollars per day per active user, and so it's, like, it does come out to a bit higher over the course of a month in Cursor but I don't think it's, like, out of band, and that's, like, roughly how we're thinking abo…”
Anthropic estimates Claude wrote 80% to 90% of the Claude Code codebase
“Probably near 80, I'd say.”
Fanelli: Series A Startups Have 80% to 85% AI-Generated Code
“We're hearing a lot of that in our portfolio companies, like, more like series A companies, it's like, 80, 85% of the code they write is ad generated.”
Cherny: Claude Code operates as both an MCP client and an MCP server
“Because Cloud Code is an MCP client and an MCP server.”
Anthropic runs Claude Code in CI/CD to semantically lint GitHub pull requests
“For quad code internally in the GitHub repo, we have this GitHub action that runs. And the GitHub action invokes quad code with a local slash command. And the slash command is lint. So it just runs a linter using quad. And it's a bunch of things that are prett…”
Cherny: Rich CLI development resembles cross-browser issues of the IE6 era
“So building in this way, it feels to me a little bit like a building for the browser back in the day where you had to think about like,
Internet Explorer six versus Oprah versus like Firefox and whatever.
Like you have to think about these cross-terminal diffe…”
Cherny: Anthropic uses Bun to compile Claude Code executables
“So we use Bunn to compile the code together.”
Anthropic: Claude Code uses regex allowlisting and default read permissions
“We're spending a lot of time building out the permission system, so Robert on our team is leading out this work. We think it's really important to give developers the control to say, hey, these are like the allowed permissions. Generally, this includes stuff l…”
Cherny: Coding agents require early failure detection over delayed intervention
“And what we find is that if the model is doing something wrong, it's better to identify that earlier and correct it earlier, and then you're gonna have a better time. If you wait for the model to just go down this like totally wrong path and then correct it 10…”
METR benchmark: AI agent autonomy duration doubles every 3 to 7 months
“They established a Moore's law for time between human input, basically, and it's basically doubling every three to seven months is the idea. And Enthopic is currently doing super well on that benchmark. It's roughly about autonomous for 15 minutes at the 50th …”
Anthropic tests Claude for internal code reviews, not yet releasing it
“We have some experiments where Quad is doing code review internally. We're not super happy with the results yet, so it's not something that we want to open up quite yet.”
Cherny: Anthropic is currently bordering on AI Safety Level 3 capabilities
“Yeah, we're kind of bordering on three right now.”
Cherny: I have not manually written a unit test in months
“So for example, like, I have not manually written a unit test in many months.”
Anthropic: Cycle time and backlog unlocking are key productivity metrics
“The two that we're really trying to nail down are, one, decrease in cycle time. So how much faster are your features shipping because you're using these tools? So that might be something like the time between first commit and when your PR is merged. It's very …”
Anthropic engineer built a Slack bot using Claude Code to automate PRs
“There was a really early version of Cloud Code many, many months ago, and this one engineer at Anthropic, Jeremy, built a bot that looked through a particular feedback channel on Slack, and he hooked it up to code to have code automatically put up PRs. With ju…”
Cherny: AI prototyping informs engineering decisions faster than writing design docs
“Before I would write a big design doc, and I would think about a problem for a long time before I would build it sometimes for some set of problems, and now I'll just ask quad code to prototype, like, three versions of it, And I'll try the feature and see whic…”
Wu: Internal operational tools are a major use case for Claude Code
“You mentioned internal tools, and that's actually a really big use case that we're seeing emerge, because a lot of times if you're working on something operationally intensive, if you can spin up a internal dashboard for it, or like an operational tool where y…”
Cherny: Foundation models will eventually subsume external memory and RAG architectures
“Everything is the model. Like that's the thing that wins in the end. And it just, as the model gets better, it's it subsumes everything else. So, you know, at some point the model will encode its own knowledge graph. It'll encode its own like KV story if you j…”
Cherny: Early Claude Code prototypes used vector RAG with Voyage AI
“Originally we tried very, very early versions of Claude actually used RAG. So we like indexed the code base and I think we were just using Voyage.”
Cherny: Agentic search trades tokens and latency for better security and accuracy
“So essentially at the cost of latency and tokens, you now have really awesome search. Without security downsides.”
Cherny: Claude works best when researching code before planning execution
“Generally, the usage pattern that works best is you ask Quad to do a little bit of research, like use some tools, pull some code into context, and then ask it to think about it. And then it can make a plan, you know, do a planning step before you execute.”
Cherny: Claude Code uses pure chain-of-thought, not Think Tool
“Yeah, this is, it is, it's all chain of thought, actually, in quad code. So we don't use the think tool. Anytime that quad code does thinking, it's all a chain of thought.”
Cherny: Mandating Docker sandboxing for AI coding creates too much developer friction
“Ideally, the thing that we want is to always run code in a Docker container, and then it has freedom and you can kind of snapshot, you know, with other kind of tools later on top, you can snapshot, rewind, do all this stuff. Unfortunately, working with a Docke…”
Cherny: Claude Code spawns parallel sub-agents to investigate complex coding tasks
“And so in the UI, when you say, when you see a task that's actually like a sub-Claud, it's a sub-agent that does this. And usually when I do something hairy, I'll ask it to just investigate, you know, three times or five times or however many times in parallel…”
Anthropic: Most of Research Team Uses Claude Code Daily
“Most of our research team actually uses quad code day to day, and so it's been a great way for them to be very hands-on and, like, experience the model failures, which makes it a lot easier for us to target these in model training and to actually provide bette…”
Wu: Claude 3.7 Sonnet is highly persistent but interprets tasks too literally
“The latest Sonnet three seven is, it's a very persistent model. It's like very, very motivated to accomplish the user's goal, but it sometimes takes the user's goal very literally, and so it doesn't always fulfill what, like, the implied parts of the request a…”
Wu: Anthropic plans native cross-session resumption for Claude Code
“We plan to build in more native ways to handle this specific workflow.”
Claude Code's markdown parser was generated by Claude in two prompts
“And so the night before the release at like, 10 PM, I'm like, all right, I'm going to do this. So I just asked Quad to write a markdown parser for me. And they wrote it.
It wasn't quite zero shot, but after, you know, like maybe like one or two prompts, it got…”
Boris Cherny: Developers will increasingly generate custom libraries using AI
“I expect that a lot more people are going to start doing things like this, like writing your own libraries or just shipping every feature.”
Cherny: Claude Code delivers up to 10x productivity gains for Anthropic engineers
“Anecdotally for me, it's probably two X my productivity. So I'm just like, I'm an engineer that codes all day, every day. For me, it's probably two X. Yeah. I think there's some engineers at Anthropic where It's probably 10 X their productivity”
Anthropic's non-coding product designer ships monorepo pull requests using Claude Code
“And she's landing PRs to our console product. So it's not even just, like, building on quad code. It's building, like, across our product suite in our monorepo.”
Cherny: Prompt Engineering Skills Determine Success With Coding Agents
“And one thing we find is that people that are really good at prompting models, From whatever context, maybe they're not even technical, but they're just really good at prompting. They're really effective at using code. And if you're not very good at prompting,…”
Cherny: Prompt engineering skills for coding may become obsolete within 3 months
“But I also agree that, you know, maybe in a month or two months or three months, you won't need this anymore because, you know, the bitter lesson always wins.”
Cherny: Claude Code is the thinnest possible wrapper over the underlying model
“All the secret sauce, it's all in the model and this is the thinnest possible wrapper over the model. We literally could not build anything more minimal. This is the most minimal thing.”
Anthropic uses Claude to rewrite Claude Code from scratch every 4 weeks
“We've rewritten it from scratch, yeah, probably every three weeks, four weeks or something, and it just like all the, it's like a ship of Theseus, right? Like every piece keeps getting swapped out, and just because quad is so good at writing its own code.”