Evan Kormos
A Coinbase engineer focused on the deployment and evaluation of AI agents in business processes, who proposes methodologies based on policies, DAG/DCG and evaluation systems.
Timeline (last 90 days)
Coinbase engineer Evan Kormos, in an AI Engineer talk, proposed a methodology for judging whether an AI agent can operate unattended, including a complete framework from flowcharts to an evaluation system and then to gradual delegation of authority.
Kormos calls policy "the organization's original intent" or "the original system prompt", and emphasizes writing it with explicit normative verbs such as must, may, and should.
Kormos views the agent as a cross-cutting operational layer in the people-process-technology triad model: it collaborates with people, executes part of the process, and is itself constituted by technology.
The core recipe of Kormos's method is "keep determinism on the outside, keep adaptivity on the inside": the outer loop—process, gates, and policy boundaries—stays stable, while the inner loop of "generate—check—iterate" allows the agent to try repeatedly.
After an agent is added to accelerate a certain step, the bottleneck keeps shifting, so the overall observability and measurement of the process is more important than the writing of the workflow tool itself.
Kormos believes the trustworthiness of a loop equals only the trustworthiness of the check; defining what counts as "good" is the hardest part and requires a reliable evaluation method.
Kormos proposes a phased delegation path, "hill climbing": full supervision, sampled inspection, exception handling—unattended operation is not a switch but a gradual curve with metrics, thresholds, and the ability to roll back.
All Facts (7)
Coinbase engineer Evan Kormos, in an AI Engineer talk, proposed a methodology for judging whether an AI agent can operate unattended, including a complete framework from flowcharts to an evaluation system and then to gradual delegation of authority.
50%UnverifiedKormos calls policy "the organization's original intent" or "the original system prompt", and emphasizes writing it with explicit normative verbs such as must, may, and should.
50%UnverifiedKormos views the agent as a cross-cutting operational layer in the people-process-technology triad model: it collaborates with people, executes part of the process, and is itself constituted by technology.
50%UnverifiedThe core recipe of Kormos's method is "keep determinism on the outside, keep adaptivity on the inside": the outer loop—process, gates, and policy boundaries—stays stable, while the inner loop of "generate—check—iterate" allows the agent to try repeatedly.
50%UnverifiedAfter an agent is added to accelerate a certain step, the bottleneck keeps shifting, so the overall observability and measurement of the process is more important than the writing of the workflow tool itself.
50%UnverifiedKormos believes the trustworthiness of a loop equals only the trustworthiness of the check; defining what counts as "good" is the hardest part and requires a reliable evaluation method.
50%UnverifiedKormos proposes a phased delegation path, "hill climbing": full supervision, sampled inspection, exception handling—unattended operation is not a switch but a gradual curve with metrics, thresholds, and the ability to roll back.
50%