AI code development as Coinbase’s new default
AI code development is the systematic use of large language models and related tools to assist or automate software creation across the lifecycle, from prototypes and internal tooling to production systems, in order to increase speed, reduce manual workload, and reassign engineers to higher-value design and review tasks.
Coinbase has not dipped its toes into AI-assisted coding; it has jumped in headfirst. Its Head of Platform, Rob Witoff, says that 95% to 100% of the company’s code is now written with large language model assistance, up from 40% in February. That is a 2.4x acceleration in AI adoption in a matter of months—a shift no serious fintech can ignore. AI is now embedded in daily development, with each engineer using around five to ten AI agents that together perform work equivalent to roughly 1,200 digital workers. This is not a pilot or an experiment. It is a deliberate bet that fintech software automation is the only way to keep startup-level speed in a heavily regulated, high-stakes environment.
From 40% to 95%: why Coinbase moved so fast
The speed of Coinbase’s shift from 40% to near-total AI code development is not an accident; it is a response to organizational pressure and strategic intent. The company cut 700 staff in May, and its CEO Brian Armstrong framed the move as a need to “return to startup speed with AI at the core”. When leadership defines AI as the way back to agility, experimentation stops being optional. AI is now deeply integrated into the firm’s daily development processes, not treated as a side tool or a novelty.
The logic is straightforward: if large language model agents can produce code equivalent to the output of about 1,200 employees, the economics of traditional headcount-heavy engineering are blown apart. Coinbase has made AI the default, keeping humans in charge of critical cryptographic algorithms and key modules while handing over early-stage work to automated systems. That trade-off shows a clear, opinionated stance: use AI for speed and scale, reserve human expertise for correctness and risk.
What LLM agents changed in Coinbase’s development model
Coinbase’s use of LLM agents is not about a single coding assistant sitting in an IDE; it is about small teams orchestrating a swarm of specialized tools. Each engineer now runs five to ten AI agents in parallel, which together deliver the coding output of around 1,200 workers. This scale of LLM assisted coding changes how work is planned. Prototype development is fully automated, according to the company, while humans step in to review core cryptographic and other critical parts of the stack.
The result is a more compact but more senior engineering team. Coinbase claims that two to three people can now do what took more than ten before. That is fintech software automation in practice: AI agents draft, refactor, and extend code; engineers curate, approve, and design. The controversial part is not the tooling; it is the redefinition of what a “team” is when machines handle most of the keystrokes. Coinbase has decided that the future engineer is a conductor of LLM systems, not a line-by-line coder.
A new playbook for fintech software automation
Coinbase’s path suggests a new playbook for enterprise AI adoption in financial services. First, identify where machines can safely take over: in this case, non-critical paths and prototype work are handed to AI agents, while humans guard the cryptographic and security-sensitive core. Second, restructure teams around AI capacity, not headcount, assuming that an engineer plus a cluster of agents represents something like a ten-person team. Third, embed AI into daily workflows instead of treating it as an optional tool—the company states that AI is now deeply woven into everyday development.
The uncomfortable truth is that this model will pressure competitors. If one fintech can run a leaner, senior-heavy team that builds faster and cheaper through LLM assisted coding, others will either adapt or explain to their boards why they did not. Coinbase’s move from 40% to nearly 100% AI-assisted code is less a curiosity and more a warning: in the next wave of fintech software automation, “manual by default” may simply be too slow.
