The bet was placed long before there was a product

Before Anthropic shipped any product, it had already bet on the idea of having AI write code on its own. The first thing they built was itself a coding assistant: a VS Code extension that took a prompt and gave you four different suggestions. By spring 2022, it had roughly 100 external users. Models were bad back then, and the infrastructure was primitive.

Shauna Kravec says the infrastructure for agentic coding (having AI write and run its own code) is far more complex than building a chatbot. The moment code execution is involved, you have to think about exactly which environment the code runs in and how to control it safely and effectively. Her comment gets at the continuity here: the hard problems many people are hitting with agents in 2026 are the exact same hard problems the team was hitting back in 2022.

clide: the internal marvel that was cool and painful to use

Before there was an official product, there was an internal tool called clide — a name a colleague, Eli Tran-Johnson, had given to an even earlier tool. It was a command-line tool where you could chat with Claude and have it edit code and do development work. Dawn Drain spent a long time teaching the model to write diffs (a representation of a code change — the most natural way programmers express edits), and clide is what grew out of that effort.

clide had a feature that could fan out 100 Claude Haiku instances to run in parallel, letting you ask a question against an entire large folder even when that folder couldn't fit into the context window (the amount of content a model can read in at once). Dawn Drain says he loved those moments of pair work, pulling out clide to answer a question on the spot, with people constantly asking him how he knew about all these useful tools.

Before moving to Labs, Adam Wolff added the last feature to clide, called clide agent. At the time clide had no bash tool, so it could only do so much. He got it to infer what you were trying to do from something you'd half-edited — a very early form of agentic behavior, the seed of something that could carry a step forward on its own. The first time it worked, he was dancing around his kitchen, unable to believe it.

But everyone complained clide was clunky and slow to start (Sid Bidasaria). Boris Cherny says you had to recite a bunch of incantations just to get it running. It wasn't great software, but it had something magical and powerful about it — it had glimpsed the future. All the puzzle pieces were there; they just hadn't been assembled yet.

A demo nobody understood, built in two days, changed everything a year later

Boris Cherny's starting assignment was just three words: "automate programming." Nobody told him what that meant in practice. He first taught himself to use the API, tinkering as he built, and eventually put together a demo he named Claude CLI (command-line tool).

Nobody understood the demo, and he hadn't fully worked it out himself. But looking back now, all the original elements are still there. He asked it what music he was listening to, and it just took a screenshot of Apple Music and read it out — that's how it pulled it off. It took roughly two days. Rebuilding the same thing with Claude Code today would take two minutes. He posted it to Slack and got two or three likes.

The next day he walked into the office and saw Robert Boyce working, his screen showing those now-iconic red-and-green lines of a code diff. "Yeah," Robert said, "it's helping me write code." That's the moment it became useful. Ben Mann had the realization on the spot: all the puzzle pieces existed — they just needed putting together. After that, Boris started working on it every single weekend; friends would invite him out, and his mind just couldn't let it go.

Sid Bidasaria says the codebase had zero pre-commit review gates — if someone said "this doesn't work well here," they'd push a fix within minutes, and users had it in their hands five minutes later. Boris had built solid auto-update and good user telemetry early on, which is what made this whole loop possible. As a command-line tool rather than a web app, there was no complex architecture to navigate around either — a simple client meant fast iteration.

Adam Wolff kept pushing back against rapid team expansion while Boris kept pushing to grow. Boris later admitted that, in hindsight, keeping the team small was actually key to their success: it forced everyone to be frugal with resources, avoid over-engineering, and lean harder on Claude itself — otherwise there was no way to move fast enough. By the time it officially launched to the public in February 2025, Claude CLI had been renamed Claude Code, a name suggested by product marketing's Alex Isken; the team liked how simple it was.

The launch also carried two little easter eggs hidden in the terminal. Late one night before launch, Igor Kofman thought it'd be great to have an ASCII art logo, so he sat down with Claude and worked through several sets of character-art fonts, producing what's now the iconic all-caps CLAUDE CODE logo that pops up on login — a little surprise the moment you log in. Meaghan Choi's proudest touch was tucking a little character called Clawd into the terminal, a figure originally created by Sam McAllister for the Claude 3.5 Sonnet launch. She says opportunities to do that kind of thing in a product don't come around often.

Ship something that only works 20-30% of the time now, so the next model generation can carry it

Feedback during the early access phase before launch was lukewarm. Cat Wu says people thought the idea was cool, but there were bugs everywhere. They shipped it publicly anyway. Behind that decision is a counterintuitive product philosophy.

Meaghan Choi says the Claude 4 series is when the moment truly belonged to them. Before that, the UX design they could actually do was limited — the model couldn't yet support the product they wanted to build. Once it could, design finally had room to work. Boris Cherny adds the other half: they also launched the subscription model. Two engines fired at once — a subscription business model, and a leap in model capability — together igniting Claude Code's takeoff.

Dawn Drain's take is blunt: Claude Code doesn't really owe clide that much. Once you cross the model-capability threshold, this form just emerges on its own.

Three people from three different worlds, and how they first got their hands on it

After launching as a research preview in February 2025, it landed in three very different settings: a big-tech company evangelizing it desk by desk, an open-source author bending his workflow around it, and a nonprofit finally able to afford custom software it could never have paid for before.

Tristan Hume's experience was the flip side. Most of what he was doing at the time required deep background knowledge — writing low-level kernels for accelerators, with a lot of details that simply weren't documented publicly anywhere. Early Claude Code wasn't great at building its own tools or learning on the fly, so for a while it wasn't much use to him — good only for narrowly scoped tasks. That only turned around once later models got stronger.

Now, nobody reviews permission requests line by line anymore

Just how much daily work has been reshaped shows up in a few concrete numbers. The share of Boris Cherny's own code produced by Claude Code kept climbing.

He remembers the day Sonnet 4 came out — he was backstage at the Code with Claude developer conference, writing code, and felt the model had gotten so much better, so much more capable of acting on its own. By winter 2025, 100% of his code was written by Claude Code — not a single line by hand. He mentions one day where he made 88 commits, with his wife and dog on the couch beside him the whole day while he coded.

Shauna Kravec now runs a whole fleet of 12 Claudes at once — reading docs, updating things, pulling messages from Slack. As head of research, she barely wrote code herself for years; now she writes more, because the barrier is lower. Igor Kofman takes it a step further: as models and Claude Code keep getting better, the next layer of abstraction isn't managing a bunch of Claudes — it's managing the managers of those Claudes.

Cat Wu offers a trust-level observation: at first, people would read every single permission request Claude Code sent, line by line; now a large share of users just auto-approve everything. That shift says Claude has earned their trust.

React taught us that anything successful eventually becomes something else

To close things out, a few key people offer judgments drawn from their own experience — all factual comparisons, no deification, no predictions.

Adam Wolff draws an analogy to React. React started as a pretty pure computer-science idea: using functional programming to represent UI state, better than message passing. By the time it hit millions of daily users, it had become something else — a logo, a brand, a feeling, far beyond the original technical insight. Most React users who love it couldn't even trace it back to that original insight. He predicts Claude Code will evolve the same way: whether you think Claude Code is the terminal, is Claude's personality, or is some prompting trick — at the limit, none of that matters.

Tristan Hume pushed it to the limit once: he had Claude build a Jupyter Notebook replacement from scratch, never looking at a single line of code, even having it test the UI itself using a browser. It genuinely produced something usable. But once he actually used it, he found he didn't like it. He says what he's waiting for is a Claude with taste — one that knows everything he needs and can just go handle it in the background.

Shauna Kravec's take draws on her background in theoretical physics. Some hard problems have taken humans hundreds of years to solve, with progress far slower, often gated by expensive, hard real-world constraints — you can only build so many particle colliders. Her view: through most of 2026 and 2027, a great deal is likely to happen within any given three-month window; a three-month stretch in 2024 also counted as progress, but nowhere near this intense. She's not sure anyone's ready for that kind of mismatch in pace.

Boris Cherny closes with a set of physical objects. He shows an IBM 029, much like the one his grandfather used in the Soviet Union; then one of the earliest text editors, still installed on every Mac today. It kept evolving, evolving, evolving, always changing. Claude Code sits somewhere on that same lineage.

The change started with something unassuming. Engineer Boris Cherny's assignment was just three words — "automate programming" — and he taught himself as he built, putting together a demo in two days that got only two or three likes when he posted it to Slack (the company's internal chat tool). The next day he walked into the office and saw a colleague actually using it to edit code — and in that moment, the scattered puzzle pieces suddenly clicked into place.

But it only worked 20-30% of the time back then — so why ship it that early?

Co-founder Ben Mann's answer is counterintuitive: precisely because the model would only keep getting stronger, the product had to ship early. Say XiaoHu wanted to build an AI coding tool, and today's model only gets it right 20-30% of the time — he wouldn't wait for the model to improve. He'd build the product now and ship it now.

After officially launching as Claude Code in February 2025, it only truly took off once the Claude 4 series caught up and the subscription model landed alongside it. Where it landed was strikingly varied too: Ramp's engineers evangelized it desk by desk, the founder of the open-source project Bun bent his workflow around it, and an Alaskan nonprofit built an app it could never have afforded before. The clearest sign of the change is how much faster the same task got — Boris's original demo makes the cleanest ruler for that:

All figures above come from the interviewees' own recollections — this is the vendor's own account and has not been independently verified.