Meet Devin: Goldman Sachs’ new AI software engineer that never sleeps

An AI called Devin just landed a job on Wall Street. 

Goldman Sachs just “hired” an AI software engineer made by the startup Cognition. 

Goldman Chief Information Officer Marco Argenti told CNBC that the company plans to “augment” its workforce with the AI tool, which will execute tasks on behalf of its more than 10,000 human software developers. “Initially, we will have hundreds of Devins [and] that might go into the thousands, depending on the use cases,” Argenti said.

Wall Street has been wading in and exploring AI in the last couple of years, but Goldman’s push to integrate autonomous AI agents might be the finance sector’s first plunge into the deep end.

A brief history of Devin

Cognition introduced the technology, which it hailed as the world’s “first AI software engineer,” last year. That tech, known as Devin, was designed to execute software engineering tasks independently, making thousands of decisions along the way. Using natural language prompts, programmers can put Devin to work doing complex tasks like building an app or finding and fixing bugs in a codebase.

Cognition describes Devin as a “a tireless, skilled teammate, equally ready to build alongside you or independently complete tasks for you to review” – a description as likely to strike fear in the hearts of workers as it is to have executives seeing dollar signs. Shortly after its launch, the founder of prominent AI search engine Perplexity praised Cognition’s Devin as the first AI agent “that seems to cross the threshold of what is human level and works reliably.”

The company was valued at $4 billion in March after raising hundreds of millions in investment led by 8VC, an early stage venture capital firm founded by Palantir co-founder Joe Lonsdale. Cognition is obviously leaning on Devin’s anthropomorphic branding to generate buzz and make Devin feel like a one to one substitute for a human that can write code, but so far that strategy seems to be working.

Like most promising new technologies, the marketing doesn’t always match reality – at least not yet. When one group of data scientists put Devin to the test, they found that the AI software engineer only successfully completed three of 20 proposed tasks, with 14 of the test projects being total flops. In benchmark tests, Devin performed twice as well as an LLM-based chatbot, but didn’t deliver on the promise of operating fully autonomously.

AI-generated code can also pose its own problems for companies, with bugs leading to downtime and security risks being glossed over due to less human involvement in the process. 

What is agentic AI?

Agentic AI is yet another AI buzzword, but it does mean something specific. Unlike chatbots or AI research tools that people are most likely to interact with, agentic AI is designed to execute tasks and make decisions on its own without constant human input (hence the emphasis on agency in “agentic”). 

The ability to execute multi-step jobs from start to finish means these tools work more how humans do, pursuing set goals and completing various kinds of tasks to get there. While generative AI is focused on generating outputs, like writing a draft email or compiling research, agentic AI is all about executing tasks and taking action, though how those systems accomplish that and the degree to which they are successful is up to their design.

Wall Street wades in

Goldman Sachs seems to be the only major bank implementing a start-to-finish AI coding tool like Devin, but its competitors likely aren’t far behind. AI assistants and chatbots are used widely on Wall Street already as big banks look for a competitive edge and pour investment into AI-related hires and technology. 

JPMorgan Chase introduced its own generative AI assistant internally last year, making the tool available to 60,000 employees who can leverage it to write emails and file reports. Morgan Stanley similarly provided financial advisors with internal tools running OpenAI’s tech under the hood.

“We are completely convinced the consequences will be extraordinary and possibly as transformational as some of the major technological inventions of the past several hundred years,” JPMorgan Chase CEO Jamie Dimon said in bullish comments on AI last year, adding that he expects the technology to “augment virtually every job.”

While AI chatbots and other lower-level tools are approaching ubiquity, adoption of agentic AI isn’t there yet. In a report last year, Deloitte predicted that a quarter of companies already using generative AI would begin exploring agentic AI in 2025, but risks from the technology’s relative lack of human oversight means companies will move more slowly to fully implement it.