# How to Know When an AI Agent Can Work Unsupervised — Evan Kormos, Coinbase

## Executive summary

This talk outlines a framework for integrating AI agents into highly regulated, policy-driven workflows, such as those found in enterprise financial services. The core methodology involves modeling processes as Directed Acyclic Graphs (DAGs) for the overall flow, and Directed Cyclic Graphs (DCGs) for internal nodes (loops). Trust in agents is established not by the model itself, but by implementing robust confidence mechanisms—specifically, rigorous evaluation (Evals) and tracing of runtime logs—to ensure outputs meet defined policy thresholds. Agents should be deployed in a phased manner: fully supervised $\rightarrow$ sampled $\rightarrow$ exceptions only.

## Key takeaways

- Policy as the System Prompt: The organizational policy serves as the 'original institutional intent' or the 'original system prompt,' defining rules using language like 'must,' 'may,' and 'should' to govern how work is done.
- The Loop is only as good as the Check: In any process loop (DCG), the reliability of the entire workflow depends entirely on the quality and rigor of the check mechanism used to determine if the work is 'good enough' to proceed.
- Phased Agent Adoption: Agent deployment should follow a structured progression: 1) Fully supervised (manual verification); 2) Sampling (humans review a slice of output); 3) Exceptions only (humans handle only failures).

## Technical details

- Workflow Modeling (DAGs/DCGs): Workflows are modeled as Directed Acyclic Graphs (DAGs) to represent the overall, gated flow (e.g., approval processes). Individual steps within the flow are often loops, represented by Directed Cyclic Graphs (DCGs). Tools like Apache Airflow manage DAGs, while LangChain/LangGraph are suggested for managing DCG state.
- Confidence Mechanisms (Evals and Traces): Since agents are probabilistic, confidence is built using statistical analysis through 'Evals' (evaluations on representative datasets) and 'Traces' (runtime logs of what happens). This complements traditional software confidence mechanisms like unit tests.
- System State Separation: It is crucial to separate the 'build state' (defining or changing policy) from the 'run state' (the system constantly operating and moving).

## Practical implications

- Operational visibility is paramount: As agents increase efficiency and parallelize processes, the focus must shift to monitoring the process flow and identifying new bottlenecks, rather than just building the workflow tool.
- To maximize agent utility, the process must be designed to be deterministic on the outside (policy boundary) but adaptive on the inside (the loop).
- When determining if an agent should own a node, assess if the node can be checked against the policy easily, run in a bounded loop, and be traceable.

## Topics

AI Agents, Workflow Automation, Compliance and Policy, Graph Theory (DAGs/DCGs), MLOps, Evan Kormos, AIE CODE summit

Source: https://www.youtube.com/watch?v=_4Ph2dhmUFI