Matt Glickman, co-founder of Genesis Computing after nearly 25 years at Goldman Sachs and an early product role at Snowflake, argues that February 2026 was an "event horizon": models became capable of complex, multi-step data engineering, so enterprises must now ask why an agent cannot do a job before hiring a person. His position by the end is that agents can outperform humans at this work if a harness makes them prove their results, a crawled context graph supplies the firm's unwritten knowledge, and the system runs inside the client's own environment. The cost lands on junior hiring, not on today's staff.
From Goldman to Snowflake. At Goldman, Glickman ran the data platform for the quants, which helped the firm manage risk through the financial crisis; afterwards everyone wanted to run on one copy of the data, a "big user problem" rather than a big-data one. Around 2014 a Snowflake founder pitched separating compute from storage in the cloud to a room of Goldman skeptics. Glickman applied for the only job listed, the company thought this was how Goldman vetted vendors, and he went on to lead product, commuting between New York and the Bay Area every other week.
Why New York, and why finance. For enterprise startups he rates New York above the Bay Area for its concentration of finance, healthcare and media. Finance and healthcare, late to the cloud, are early to AI because of their operational complexity and because models are trained on so much text about them; media and gaming, early to cloud, now hang back for fear of content leaking into models.
The founding insight. Snowflake customers' AI self-service demos plateaued at about 80%, lacking both model capability and harnesses. Glickman also argues that products built for revenue beat internal tools built for cost: on Goldman's private cloud a machine took about 3 months to provision, and one quant's deleted mount point "takes out half of Goldman", whereas EC2 is honed because downtime costs money.
Proof, confidence and escalation. Glickman says humans get too much credit for accuracy, but warns against "single shot and yolo". Models do the minimum, so Genesis's blueprints set runbooks for extraction and source-to-target mapping, require artifacts proving that every one of 30,000 reports was run, and have a second AI check the results. Since reasoning models arrived "early last year", agents can state their confidence, for example 90% with one open question. Genesis inverts the co-pilot pattern: agents work for hours and escalate only when unsure, and each answer is stored for next time. Customers are now asking for work further up the stack, such as board presentations.
The living context graph. Much institutional knowledge is never documented and leaves with burnt-out data engineers. Genesis's onboarding crawls databases, repositories and SaaS tools into a graph of data relationships; given tools to navigate it, the head agent, Eve, traces a report's lineage to source a new column. Glickman predicts firm-owned graphs like this will be valued in M&A. Genesis deploys inside the client's environment, managed by an AI engineer, and the knowledge stays the client's asset.
Jobs and timing. He says existing staff will not lose jobs but junior hiring will shrink, with a few "conductors" running many agents as special hires, and he worries that schools still discourage AI. On when to jump into a shift, his test is "why am I seeing this before other people?", with windows now shorter. He notes that a famous computer scientist of about 90 had just published a paper co-written with an AI, but says AI has not yet broadly discovered new things.