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

Highlights

• Software policy operates as institutional intent or the original system prompt, using terms like must, may, and should to define organizational rules.

• Directed acyclic graphs, or DAGs, represent workflows and their dependencies as steps that move forward without looping back.

• Directed cyclic graphs, or DCGs, manage iterative loops inside process nodes where work is reviewed and refined until it passes.

• Evaluating probabilistic AI outputs requires representative datasets, volume-based statistical analysis, and runtime traces instead of simple exact output comparisons.

• Transitioning an agent from supervised operation to unsupervised work involves clearing measured confidence thresholds and moving toward manual sampling and exception handling.

Timeline

1. 00:00 - 03:12

   Defining Policy and Operational Models

   Software development and financial services depend on baseline policies that act as institutional intent. While policies stay relatively fixed, running systems change continuously. Agents function within this ecosystem by interacting with human teams, executing parts of workflows, and utilizing underlying technology.

   • Organizational policy sets compliance rules using strict terms like must, may, and should.
   • System build states and run states remain separated while products evolve rapidly under policy boundaries.
   • AI agents integrate across people, processes, and technology as a cross-cutting operational model.
2. 03:12 - 08:09

   Structuring Workflows with DAGs and DCGs

   Workflows use directed acyclic graphs to manage dependencies for tasks like commercial lending, token listings, and software delivery. Inside individual nodes, iterative loops act as directed cyclic graphs managed by tools like LangChain or LangGraph. Introducing agents into these loops accelerates output generation, which subsequently increases pressure on adjacent steps like design and security scanning.

   • Directed acyclic graphs organize linear workflows with gated directions and parallel reviews.
   • Directed cyclic graphs manage internal node loops where work iterates based on feedback until it meets quality standards.
   • Adding agents to workflows shifts volume and creates new bottlenecks that require operational visibility.
3. 08:09 - 13:12

   Building Confidence Through Evaluations

   AI engineering faces challenges due to input and output variability. Injecting DAG state information helps agents understand their broader context. Because generative outputs are probabilistic, confidence relies on systematic evaluations across representative datasets and volume-based statistical analysis rather than isolated test checks.

   • High-quality inputs and injected graph states improve agent output consistency and reduce compute costs.
   • Traditional exact output comparisons fail for probabilistic agents that produce multiple valid answers.
   • Statistical evaluations combined with runtime traces form reliable confidence mechanisms.
4. 13:12 - 17:23

   Scaling From Supervised to Unsupervised Agents

   Determining whether an agent can run unsupervised depends on specific workflow criteria, such as whether performance can be easily checked against policy. Organizations scale agent autonomy by starting with day-one supervision, moving to randomized human sampling of outputs, and ultimately restricting human intervention to edge-case exceptions. Quality remains a collaborative effort driven by AI engineers and domain experts.

   • Workflow nodes are evaluated for automation potential based on policy checkability, bounded loops, traceability, and human fallback options.
   • Deployment scales from full supervision to human sampling and finally exception handling once confidence thresholds clear.
   • Deterministic outer wrappers combined with adaptive inner loops create successful agent integration recipes.