Architecture & Patterns: Single Agent Workflows

Production Patterns

Architecture View

graph LR
  A[Question] --> B[Plan steps]
  B --> C[Execute step]
  C --> D[Observe result]
  D --> E[Stop or continue]

Common Failure Modes

Common Failure Modes

• No explicit stop policy.
• Finalizing answers before enough evidence is gathered.
• Ignoring low-information repeated observations.

Interview Q&A

Q: Interview Q: What is a practical evidence threshold for a research agent?

Use a minimum number of independent observations plus relevance checks against the question scope. Thresholds should vary by task risk.

Q: Interview Q: Why log stop reasons?

Stop reasons explain behavior and speed debugging. They also help tune loop policies using real run data.