When an AI agent can work unsupervised, Coinbase's Kormos says

Kormos said he has spent 20 years delivering, designing, and building software, the last 15 in enterprise financial services. He is currently a player-coach at Coinbase.

He framed the talk around trusting agents in policy-driven workflows.

Kormos said the starting point is policy, which sets the rules for how work is done, especially for risk and compliance.

He described policy as the original institutional intent, or the original system prompt, ideally written down with language such as must, may, and should.

Kormos separated defining or changing policy, which happens at discrete moments, from the running system, which is constantly moving and changing.

Everything is still expected to stay roughly bound by policy even as product-driven changes evolve quickly.

Kormos said agents cut across the familiar model of people, process, and technology.

He said agents work with people, execute parts of the process themselves, and are made up of technology, and that the talk would focus on where agents meet process.

Kormos said a change may need approval before going live, or a customer may need verification before an account is opened, so workflows are worth mapping out.

He described a directed acyclic graph as a way to represent steps and dependencies in a flow that moves forward without looping back.

Independent reviews can now run in parallel and be brought together while keeping required checks in place, which he said was not practical before agents.

Kormos said agent adoption has been high in software delivery, and the aim is to bring the same efficiency and scale to other parts of the business.

He used software delivery to examine what happens inside a single node of a process, which he said is a loop.

Kormos described a code review node: work passes to the next step only once it is judged good enough. Often it is not, so feedback drives repeated iteration.

He said this makes the workflow a directed cyclic graph, and that LangChain LangGraph can help manage its state. He added that the graph shown is an oversimplification and real ones can be much more complex.

Kormos said exception flows are needed for cases such as code that is canceled or fails security standards, which teams should know about without designing for every edge case.

He said that once agents join, volume shifts across the flow: faster code review forces faster design or security scanning.

He argued that operational visibility across the process matters more than building the workflow tool itself, because added parallelism creates new bottlenecks.

Kormos said he believes higher-quality inputs produce better outputs at lower compute cost, though he framed inputs mainly as context and model.

He said injecting the DAG state into an agent tells it where it fits in the larger picture.

Kormos said traditional software builds confidence mechanisms such as test cases and unit tests, run as part of the delivery flow so code changes do not break other things.

Kormos said the same request can produce different answers, and more than one answer may be acceptable, so evidence is still needed but exact output comparison alone is not good enough.

First, DAG and DCG are basic graph theory; the work is in building existing processes and bringing agents into them. The recipe, he said, is to keep the process deterministic on the outside and adaptive on the inside.