Replit's CTO: Our Plan to Beat Anthropic and OpenAI

The audio and video are in Spanish. These notes are a translation.

Luis Héctor Chávez is the CTO of Replit, the platform for building software with AI that more than 30 million people use and that is worth $9 billion today. He is Mexican, and he runs all of engineering at a company fighting OpenAI and Anthropic for the future of software.

This is his second time on the podcast. We recorded exactly a year ago, and the way he describes that stretch is that he feels he has been inside five different companies within Replit.

His most direct thesis: the differentiation is not in the model, it is in the harness. And his ambition is not to solve the technical side of the business for you. It is to solve the whole business.

Structure and Key Highlights

A Year that Felt Like Five Companies

• Every time they benchmark against other companies they need a separate segmentation, because they do not compare with what a traditional company at that stage would be dealing with
• Software engineering processes are falling apart: the sequence of phases can no longer be followed
• His chronology of the problem: a month ago it was that there was so much code they could not do good code reviews. Now nobody even reads the code
• The current state: there are so many changes that they do not even know what is happening
• His expectation: in two or three months the conversation will be a different one
• What survives from the old way: certain industrial engineering processes still shape how they think. Everything else they are throwing out

On Whether the Speed Is Sustainable

• His favorite analogy: installing solar panels, where every year the organization that keeps the statistics projects the end of the exponential phase and misses by an order of magnitude
• He sees the same thing in the capability curves of the models
• The pattern he observes: every time they say a new breakthrough is needed, one shows up with the next model

Different Planning Horizons by Team

• Infrastructure has roadmaps of six to nine months, because if the VMs do not work nothing works
• What they are shipping today started in September or October of last year
• Enterprise plans by quarters, often tied to what a partner delivers
• Product cannot plan more than two or three months out
• Models: there are two weeks of rumors and two or three real days to adjust the harness before it ships

What Got Heavier

• The volume of simultaneous change, with more engineers and more agents per engineer
• What used to take months now takes weeks, with proofs of concept in days
• And they are not patches: they are complete reimplementations from scratch, three attempts in a row if needed
• The radiologist analogy Enzo brings in: AI took the easy cases and the radiologists burn out because only the hard ones are left
• His confirmation: AI fully solves a junior engineer’s work if you break it down for it, and when you have to get involved it is because you hit the limit

What the Models Are Missing

• They lack taste. They are very good at paint-by-numbers when you give them the outline
• With ambiguous tasks they find a way to do them, though not in the most elegant or efficient way
• They work better the better the objectives, the guidelines, and the feedback tools are
• Where they break: the subjective. His example, how do you quantify that something is ergonomic
• A concrete compliment to Fable: it is good at catching odd subtleties and deciding whether they were intentional or whether they are bugs

Why They Bet on the Harness and Not the Model

• They use a mix of models and continuously evaluate which one for each objective
• All of their innovation is in the harness: that the templates play well with the model’s priors and that the tools work inside Replit
• The concept they are chasing: the pit of success, steering the model so hard that it has to leave the lane by a wide margin to get it wrong
• How they move as the models get better: upward through the layers of abstraction
• The before and after: before, they watched for common vulnerabilities like SQL injection. Today they solve for the business logic not allowing improper access

Expensive Model Against Cheap Model

• The apparent tradeoff is speed and money against quality
• What they observed: when an error shows up, the agents accumulate it. You do not have one error, you have ten or fifteen
• The counterintuitive result: Opus beat Sonnet on total cost, because it solved the problem in one pass instead of ten
• The nuance: if your use case is very specific the equation changes, and general-purpose programming is harder to optimize
• The comparison with legal work: there you have case law and clearer criteria. In code, translating business requirements has hundreds of thousands of valid options

Solve the Business, Not Just the Code

• From the start the aim was to deliver the complete application: deployment, security, domain, TLS
• Where it is heading: not that you show up with a business idea and Replit solves the technical part, but that it solves the whole business
• The list he enumerates: create the application, the marketing content, the videos for social, the flywheel, run the experiments, handle maintenance
• On the idea of graduating to other tools: they do not want that perception to exist, and they already have clients at millions of ARR who did not feel the need
• When there are specific legal or technical requirements, they adapt the platform instead of letting the client go

The Agent Identity and Permissions Problem

• Companies have mature answers on data governance: who accesses what, with a policy in some system
• With agents that is close to impossible, and current technologies are not working
• There are five or six proposals for giving agents an identity and none of them is being adopted at the speed that is needed
• The underlying difference Enzo points out: an employee has a contract and consequences, and an agent has no accountability
• The problem with open standards like OAuth: the permissions are too broad. You can write every email or none of them, with no nuance
• What Replit wants to do differently: have an authorization model evaluate the alignment between the action, the written policy, and what the human asked for
• The practical goal: reduce how many times you have to interrupt the user to ask for permission
• What they are building: a full rearchitecture with a more granular authorization layer, one that respects the SSO, SCIM, and IDP the client already has, so they do not create shadow IT

Agents Already Run All Day

• His correction to the question: it is not that we are heading there. It is already happening
• What is common today: someone tells you they have 300 agents running, without knowing how many are useful
• At Replit they already have feedback loops that close on their own internally, as an experiment to turn that into a product
• Which cases justify the spend: optimization problems with a measurable objective. SEO, A/B tests, social campaigns
• His way of bounding it: you give it an objective and a spend limit, and it runs until it hits the objective

Who Buys Inside the Company

• Unlike traditional software, this cuts across the company and does not belong only to the CIO
• Sometimes IT shows up with no budget, sometimes product with no security criterion, sometimes every department at once
• His joint reading with Enzo: companies are still figuring out how artificial intelligence gets bought
• Replit’s distinctive customer is not the developer: it is product, operations, marketing, and data, who used to depend on the CTO for internal tools
• What they want for this year in marketing: that you can execute your ideas without permissions, without gatekeepers or a prioritization queue

Why Git Is the Right Primitive

• When they launched Agent 4 they modeled tasks using Git branches, letting it clone itself between 4 and 64 times depending on the plan
• The underlying reason: the models were trained on repositories, so anything that does not look like one throws them off
• Replit does not model a repository. It is a repository, and they only added forking, cloning, and pushing
• The experience Enzo recounts: his agent in a repo has more autonomy with half-finished instructions, while in Notion he had to repeat everything to it
• The problem coming: GitHub has scalability problems, because Git was designed for the Linux kernel and today that is exceeded by orders of magnitude
• And that design assumed humans, not agents, with an immense write rate
• His observation: that week four or five different bets came out, a sign that a primitive is missing

The Software Factory

• His way of thinking about programming from before Replit: a constraint satisfaction problem, where each constraint splits the solution space
• Why it fits now: it also helps agents to be able to measure whether they are inside or outside that bounding box
• The difference between telling it a rule and giving it a mechanism to measure the rule
• Why languages like Rust help: the compiler gives the mathematical proof of the error and moves verification earlier in the process
• What changed in the engineer’s work: before, you sketched and thought about the business logic afterward. Now you have to think it through ahead of time
• His principle: guessing is bad, and in many cases the agent should be asking
• What you learn with practice: which requirements can be underspecified and which have to be specified in detail
• The highest-leverage concept: give the model an oracle so it can verify its own work
• His examples: Anthropic replacing GCC with a clear test set, or looking for vulnerabilities where the capture the flag flag exists

On the 100X Engineer

• His nuance on the power law: in a population of engineers, the most prolific one sits an order of magnitude above the second
• What AI does is accelerate any process, multiplying that factor
• What he already sees: the engineers who were good and who also use AI well are even better
• Where he expects the same thing outside engineering: mathematics, optimization, and bioengineering
• Enzo’s reading: what multiplies is not doing more. It is designing the feedback loop, and for that you need the evaluations

What He Expects in 12 Months

• He is almost sure there will be another breakthrough, without knowing whether it will be in capabilities or in cost
• What seems likely to him: that they saturate one of the current benchmarks, reaching 100%
• What that implies: having to invent a new metric
• What keeps surprising him: not the advance itself, which he already expected, but feeling it
• On whether he will stop programming: what drew him to it is still intact, being able to express an abstract idea and make it real. The mechanism changed, not the patterns

Silicon Valley Glossary

People, Companies and References

People Mentioned

• Luis Héctor Chávez: CTO of Replit
• Amjad Masad: CEO of Replit
• Greg Brockman: cited for the line that the model is no longer the product
• Andrej Karpathy: for his idea of the 100X engineer
• Enzo Cavalie: host of the podcast

Companies and Organizations

• Replit: 30 million users, a valuation of $9 billion
• Anthropic and OpenAI: the competitors with their own harnesses, Claude Code and Codex
• Lovable, Factory: the rest of the coding agents field
• GitHub: the primitive that is falling short for agents
• Supabase: the example that many choices converge on Postgres
• Micron: the supply chain stock Enzo bought and sold early
• Perplexity: cited for the argument of being an orchestrator and not a model

Data and References

• 30 million users and a $9 billion valuation
• Between 4 and 64 clones of an agent depending on the paid plan
• 300 agents running: the figure companies keep repeating to him without knowing how many are useful