Coinbase engineering lead Rob Witoff said this week that AI now writes between 95 and 100 percent of the company’s code, a sharp acceleration from roughly 40 percent in February 2026. Witoff framed the output as equivalent to the work of about 1,200 employees, which he called “digital workers.” The disclosure matters beyond the headline number because Coinbase is simultaneously building the infrastructure for AI agent payments on Base, including Coinbase Agent Payments, which provides MPC-managed wallets and ERC-7677 interfaces for autonomous USDC settlement. The same organization pushing agent-based commerce is now relying on agents to write its software.
The Numbers and What They Mean
The jump from 40 percent AI-generated code in February to 95 to 100 percent five months later is aggressive by any standard. Witoff’s framing of 1,200 digital workers provides a rough proxy: if the average Coinbase engineer produces a known quantity of reviewed, merged code, the AI output represents a significant multiple of that baseline. But the metric itself is imprecise. Code generation measured by line count, file count, or commit count can overstate functional contribution, especially if human developers are heavily editing AI output before it ships. The number tells us directionally that AI-assisted development has moved from augmentation to near-total reliance, but it does not tell us how much of that code is novel versus boilerplate, test scaffolding, or generated documentation. What is clear is that Coinbase has restructured its engineering workflow around AI generation as the default, not the exception.
Recursive Buildout of Agent Payment Infrastructure
The more consequential angle is structural. Coinbase has spent the last year building Base as a chain optimized for stablecoin payments and AI agent commerce, launching tools like Coinbase Agent Payments and participating in the x402 Foundation governance body. If 95 to 100 percent of the code powering these systems is AI-generated, then the quality, security, and reliability of that code directly affects the payment infrastructure that autonomous agents will settle transactions on. This is not hypothetical. Agent payment systems require deterministic execution paths, correct wallet abstraction logic, and reliable fee handling. A bug in AI-generated smart contract wrappers or MPC key management code could produce settlement failures or fund losses that are difficult to attribute, especially when the code itself was produced by a model whose training data and reasoning trace are opaque. The recursive dynamic is worth naming clearly: agents are building the infrastructure that agents will transact on.
Security and Audit Implications
Payment infrastructure has historically required the highest bar for code quality in crypto, with formal verification, third-party audits, and bug bounty programs serving as standard practice. AI-generated code does not automatically bypass these requirements, but it changes the review calculus. If engineers are reviewing AI output rather than writing from scratch, the question is whether pattern-matching review catches the same class of vulnerabilities that original authorship and subsequent review would surface. Empirical evidence on this is thin. The crypto industry has already seen exploits in human-written code that passed multiple audits, so the baseline is not reassuring. What changes with AI generation is scale: more code is being produced faster, which means more surface area to review, and the review process itself may increasingly rely on AI tools, creating a closed loop. Coinbase has not publicly detailed how its audit and review processes have adapted to near-total AI code generation, and that gap is worth watching.
What This Signals for the Sector
Coinbase is not alone in this trajectory, but it is the most prominent stablecoin payments company to quantify the shift publicly. If the approach produces shipping infrastructure that works reliably at scale, it will normalize AI-generated code across the stablecoin and payments stack. If it produces failures, the industry will have to confront whether autonomous code generation is appropriate for systems handling real monetary value. Either outcome will shape how other issuers, chains, and payment processors staff their engineering organizations. The 1,200 digital workers framing also has labor implications that the industry has barely begun to discuss. For now, the concrete signal is that one of the largest US-based crypto and payments companies is building its agent commerce future on code that is almost entirely machine-written, and the market has not yet reckoned with what that means for the trust assumptions underlying stablecoin settlement.
Sources
- Coinbase Says AI Now Writes 95–100% of Its Code: The Math Behind ‘1,200 Digital Workers’ — Bitcoin News
- Coinbase’s Jesse Pollak hands Base app leadership to Cobie after admitting social bets fell short — The Block
- Base’s Jesse Pollak Admits Creator Coin Strategy Failed, Signals AI Pivot — The Crypto Times