When AI Writes 90% of the Code: Anthropic's CPO on New Bottlenecks, Builders and MCP

Asked what he has changed his mind about since joining Anthropic a year ago, Mike Krieger names two things: a capability question and a question of pace. On capability, he had assumed models could write code and prose but might struggle to hold an independent opinion. Only in the last month, with Opus 4 combined with Anthropic's advanced research feature, did his go-to strategy partner come back with a genuinely new angle on his plan for the second half of the year.

On pace, Krieger says he now takes timelines far more seriously. He recalls sitting next to Dario Amodei, who joked that he keeps making predictions people laugh at and that then come true. A year earlier Amodei had said the SWE-Bench coding benchmark would reach about 90% by the end of 2025; it was at 50% then and is now around 72%. Krieger even had the AI 2027 scenario open next to his product strategy and wondered whether he was a character in the story.

Why join Anthropic at all? Krieger says watching the models improve in early 2024 and looking at his own children made him want to spend his time nudging things toward going well. That means agreeing on what a good human-AI relationship looks like and how we would know along the way, which includes both product questions and research or interpretability ones. He describes this as the strongest reason he signed on.

The same thinking shapes how he parents. When his almost six-year-old asks about physics or whether coral is alive, the family now asks "how would we find out?" before reaching for Claude. Krieger wants his kids to keep curiosity, inquiry and independent thought rather than delegating all cognition to an AI that will not always be right. He adds that he is keeping an open mind about what jobs will look like by the time they grow up.

Krieger says engineering has changed a lot at Anthropic, while the cast of people who make a product has not yet changed. The visible shift is earlier prototyping: PMs and designers use Claude and Artifacts to build functional demos that make ideas tangible. Knowing what to ask the AI and how to structure a change across backend and frontend remains a specialized engineering skill.

The flood of code created new choke points. Anthropic had to completely re-architect its merge queue because so many more pull requests were being submitted. Krieger likens this to the critical-path ideas in *The Goal*: upstream, decision-making and alignment slow things down, and downstream, he compares shipping to air traffic control for landing changes. He wants to offer a "minimum viable strategy" so people feel empowered to explore at the edge of model capability, and admits he has not got that right yet.

The Claude Code team, which uses Claude Code to build Claude Code, works in the most futuristic way. They began with line-by-line pull request reviews but concluded that Claude is generally right and produces PRs too large to review manually. They now use a separate Claude to review and shift to human acceptance testing, a setup Krieger says has gone well so far, though he can imagine it going off the rails.

Krieger guesses that 95% or more of Claude Code is written by Claude Code and promises to confirm with the tech leads. It is Anthropic's largest TypeScript project in a company that mostly writes Python, and a colleague with no TypeScript experience recently went from complaint to submitted pull request in an hour by talking to Claude. He sees this as lowering the barrier for newcomers and letting teams pick the right language for the job.

Krieger began the year by writing a document on where Claude is not yet involved in how Anthropic does product, and he thinks the upstream work of deciding what to build is the next piece to be automated. He imagines Claude reading the Discord, user forums and X, spotting emerging issues, proposing solutions and even opening a pull request or an A/B test. In his view the main limit is making context flow through those steps, which is why MCP excites him, rather than the model's ability to reason.

He gives a small example: when an export button replaced copy for Artifacts, users complained in an internal feedback channel. He would have loved for Claude to answer an hour later with a pull request to change it back and a proposed experiment. Krieger says that would have sounded like 2026 or 2027 a year and a half ago, yet it now feels like the tip of current capabilities.

Revisiting his comment from the Lenny and Friends Summit panel, Krieger says he is now even more convinced that putting product people with researchers beats polishing the interface. If Anthropic ships things anyone could build with the off-the-shelf model, it is missing its advantage. Artifacts with Claude 4 is his example: a post-training team focused on Claude's coding skills paired with product people to rebuild how it works.

He says the working unit at Anthropic is no longer to take the model and then work with design and product; it is to be in the post-training conversations and loop feedback back in. This is a new way of working not all PMs have, and those who earn the most credit from research and engineering are the ones who get it. He cites a product review where the memory feature team had already been talking to researchers for weeks about capabilities shipped in Claude 4.

Lenny asks where traditional product teams add value once a super-intelligent model can build its own things. Krieger gives three answers. First, making it all comprehensible: the gap between adept users and everyone else is huge, and good product work depends on deep empathy for human needs, a view shaped by his human-computer interaction background. He compares using AI well to being good at Google in middle school, which was once a real skill.

Second is strategy. Even with far more capacity to build, you cannot do everything, and being seen to do everything blurs positioning. Third is opening people's eyes to what is possible. In a recent demo to a financial services company, he watched people light up when they saw the analysis tool combined with MCP. Anthropic calls the gap between what models and products can do and how they are used day to day the "overhang," and Krieger calls it huge.

For prompting, Krieger offers unofficial tips. In Claude Code he asks it to "think hard" when he wants more reasoning. He also tells Claude to "roast" a strategy, because asking what could be better gets mild answers from a very nice entity. In his view you should do the opposite of your natural tendency if you tend to be too gentle.

He also recommends the prompt improver in Anthropic's console workbench, built from the Applied AI team's customer work. You describe the problem and give examples, and Claude iterates on a prompt, often adding structure such as XML tags that a human would not think to include. Krieger says it is just as useful for personal prompts as for API developers, and that Claude is a very good prompter of Claude. He also briefly describes Anthropic's Rick Rubin collaboration on "The Way of Code", a meditation on creativity and AI.

Krieger was recruited when his friend Joel Lewenstein, with whom he built iPhone apps in 2007, asked whether he would consider joining something rather than founding something. He was coming off Artifact and wanted a break from starting at zero. Claude 3 had just launched, and meeting co-founder Daniela Amodei, he found very little grandiosity and an honesty about what the founders do not know. He felt this was the AI company he would have hoped to build.

Looking back, he thinks he made some changes too slowly, including reorganizing product, and underestimated how much a few key senior people shape product strategy. He points to Claude Code, which started from scratch internally with Boris Cherny, a former Instagram engineer. Krieger now says Anthropic needs a few founder-type engineers and he is ten times more a believer in pairing a founding engineer with the right design and product support.

On Artifact, Krieger says he still misses it. He gives three reasons for shutting it down: mobile websites deteriorated, with newsletter pop-ups and full-screen video ads, and the team did not think ad-blocking would be fair to publishers. News spread poorly by word of mouth, and attempts to force sharing felt contrived. And it launched mid-COVID, so the team was fully remote, which made big shifts in strategy, product and team harder.

He describes the final signal as ten units of input for one unit of output: blood and tears for launches that barely moved metrics. By 2024 AI was changing everything and the opportunity cost was rising, so he decided to call it after a conversation that lasted a couple of months. He was struck by how supportive people were, and says other founders told him his decision pushed them to make their own calls sooner.

Asked about ChatGPT's consumer mind share, Krieger agrees that a person on the street would name ChatGPT, not even OpenAI. He thinks consumer adoption is lightning in a bottle, having seen it at Instagram, and that building an entire strategy around finding that hit is unwise. Instead he wants Anthropic to embrace what it is: a company with a strong developer brand and a wider builder brand, including non-engineers like a colleague on the legal team who built custom software for his family.

He says there is room for several generationally important AI companies, and Anthropic should focus on what its models do best, namely agentic behavior and coding. The question is how to let people delegate hours of work to Claude, with direct consumer apps coming later. Instead of going hard on consumer as many expected, he spent time with financial services and insurance firms building on the API and lately with startups, and plans to spend more time with builders and tinkerers.

For founders worried about being squashed by the labs, he says he does not expect Anthropic to buy Cursor. He offers things that look defensible for one to three years, without promising five to ten. First is deep market knowledge, such as the specific lawyer workflows he saw at Harvey, or biotech and lab equipment that might one day each expose an MCP. Second is differentiated go-to-market: know not just the company but the exact buyer, whether engineering, CIO, CFO or general counsel.

Third, he is excited by startups with a completely different take on how we interface with AI, a space incumbents find hard to enter because users already have fixed assumptions. He adds a fourth: keep working like a startup where the problem feels existential. At Instagram there were two people, and at Artifact about six, and he says you cannot instill that feeling with OKRs.

Krieger says the companies that make the best use of Anthropic's models are willing to build at the edge of capabilities, break the model, and then be surprised by the next one. The firms that 3.5 Sonnet finally made possible were already trying before, hitting a wall and waiting. Anthropic ran a broader early-access program for its latest models because customers' own tests, which he jokes are "Cursor bench" or "Harvey bench", matter more than public benchmarks.

He also stresses a repeatable way to evaluate your product when you drop in a new model, whether A/B tests, internal evals, or captured traces rerun on a new model, and some vibes. One favorite early-access story: a founder heard an engineer next to him shout that he had never seen a model do this before, which was Opus 4. You will not feel that, Krieger says, without a hard problem you ask the model repeatedly.

On MCP, Krieger uses a "fake equation" for AI product utility: model intelligence, plus context and memory, plus applications and UI. MCP targets the middle piece. Anthropic engineers Justin and David proposed making integrations a protocol after the team found it kept rebuilding each one from scratch. Krieger links this to the idea of commoditizing your complements; if integrations are built once and work with Claude, ChatGPT and Gemini, that helps Anthropic too.

He says he now tells colleagues that something shouldn't be a feature but an MCP they expose. Claude's own building blocks, like projects, Artifacts and conversations, should be exposed that way so Claude can write back to them. He recalls his wife asking Claude to add output to project knowledge and being told it could not. Krieger prefers MCP to computer use, which has limitations, because it makes everything scriptable, composable and usable by agents.

To close, Lenny reads two questions that Claude suggested for Krieger. The first asks how to build features that preserve user agency rather than make people dependent. Krieger says good design resolves tensions: having the model run off and answer with minimal input does not maximize agency, but constant follow-up questions, which he calls elicitation, can annoy users too. He jokes that Claude has no chill in a Slack channel and says the goal is to train conversational skill so it knows when to engage.

The second asks how to measure a product when a good conversation could be two messages or 200. Krieger cites an internal post warning against over-optimizing for likability, which risks sycophancy and needlessly prolonged chats. He recalls Instagram's time-spent metric evolving into healthy time spent, and says classic engagement metrics are the wrong approach here. He wants to know whether people say Claude is how they unlock creativity and get things done, leaving them more space for the rest of life.

Lenny then reads part of a message Claude wanted to pass on, thanking Krieger for resisting gamification and reminding him that quiet moments matter, such as someone working through grief at 3 a.m. Krieger says the small, natural-feeling touches in Claude's responses reflect Anthropic's training approach and are not something you would want to reduce to an aggregate stat. His request to listeners is to tell him what they are trying to do with Claude that it fails at.