▶ Watch ↗17:32
Conference affiliation: Coinbase
Evan Kormos is a software engineering leader whose work connects financial-services processes with AI agent automation. At AI Engineer NYC 2026, he was an engineering manager at Coinbase, leading teams responsible for full-stack delivery and service management. Their focus included bringing agents into policy-governed work involving Token Listings and the Coinbase Developer Platform.
Evan’s background spans enterprise transformation, retail and commercial banking, and digital assets. In his talk on unsupervised agents, he described twenty years designing and delivering software, including fifteen in enterprise financial services. His approach to autonomy starts with the business process: which work can an agent take over, how will its output be checked, and when must a person intervene?
Evan treats policy as institutional intent: written requirements, including risk and compliance obligations, that govern how an organization operates. He distinguishes policy and changes to the system from its continuously changing operation. An agent’s actions must remain within those requirements even as products and workloads evolve.
He models the overall workflow as a directed acyclic graph: steps move forward through required gates and dependencies. Inside a step, work can loop through attempts, checks, and revisions. In his code-review example, feedback prompts another revision until the work qualifies to advance. This separates the freedom to iterate from permission to move to the next stage.
The same distinction allows parallel work without removing required approvals. In an illustrative lending process, identity checks and financial analysis can proceed alongside each other, while credit review, documentation, and decisions retain their dependencies. Evan presents these as ways to design agent-assisted processes, rather than as accounts of systems he personally shipped. He also distinguishes assistance from ownership: agents may help throughout a lending workflow while the lending decision remains human-driven.
For Evan, the loop is only as trustworthy as its check. More attempts do not necessarily improve an answer. Because agents can produce different acceptable outputs for the same request, exact-output comparisons alone cannot establish quality. Domain experts must help define acceptable outcomes, and representative datasets must be evaluated repeatedly to assess performance across many runs. Execution traces explain what happened during an attempt; evaluations assess whether the resulting work meets the standard. Together, they complement conventional software tests.
His criteria for assigning work and reducing supervision follow from that approach:
Evan gives particular weight to what happens outside the automated step. Faster code review can shift pressure onto design or security checks; accelerating one stage can simply move the queue. He therefore places operational visibility ahead of workflow authoring: teams need process metrics and exception handling to see where work stalls, fails a security requirement, or needs human intervention. The measure of progress is how the whole process performs as agents take on more of its work.
▶ Watch ↗17:32
Loading more talks and highlights…