August 29, 2026
According to eFinancialCareers, Marco Argenti, Goldman Sachs' head of engineering, has found that the hard part of deploying AI agents is not getting them to write good code. General-purpose agentic systems like Claude and Devin arrive competent. What they lack is the "tricks and tribal knowledge" specific to the firm: the data standards, the security protocols, and the "engineering tenets" that make a new service fit the existing architecture and stay maintainable after the person who commissioned it has moved on. As Argenti puts it, "The unwritten rules are the ones that are actually harder to capture."
Goldman's answer is to try to write them down. The firm is interviewing its best engineers and analyzing their output to distill new "skills" that can be built into the agents. In the meantime, developers are spending a growing share of the day talking to the software in a way that would be called mentoring if there were a person on the other end. The irony is that further up the bank, the people who would normally be doing the mentoring are on the receiving end of it, with managing directors being taught to use AI by analysts and interns, which is the one context in banking where being twenty-two is an advantage. They will presumably hit the same tribal-knowledge wall once they start delegating to agents themselves.
The bigger read is what happens to judgment. Goldman partner Chris Churchman has raised the concern that bankers who spend their days fine-tuning a system and relying on its output will gradually lose the analytical instinct that made them valuable in the first place. That is a direct threat to the apprenticeship model the industry has leaned on for decades. It is also worth noting that giving juniors role models to learn from in person was one of the last arguments left for ending remote work, and that argument gets harder to make when the most important mentoring relationship an MD has is with an interface on their phone.
This matters for recruiting in four specific ways. First, the part of the job being automated is output and the part being defended is judgment, which is exactly what a technical interview tests when it asks why a number moves rather than what the number is. Second, tribal knowledge is the interview: the bar is not whether you can build a DCF, it is whether you know which assumptions inside it are actually load-bearing, because that is the layer Goldman is finding hardest to hand off. Third, reverse mentoring is real short-term leverage, and being the analyst who is genuinely fluent with these tools gets you into rooms early, but it is a door rather than a moat, because everyone will be fluent soon enough. Fourth, if agents absorb the work juniors used to learn from, the apprenticeship has to be sought out deliberately rather than arriving by osmosis. The candidates this hurts are the ones who can produce an answer but not defend it. The ones it helps were going to learn why it works anyway.
Source: Daniel Davies, "Morning Coffee: Goldman Sachs AI agents are difficult to turn into Goldman Sachs bankers. How to behave when there's 2,200 redundancies," eFinancialCareers, August 26, 2026.