Enterprise AI Measurement Guide
Agent Effectiveness
Session Depth
How deeply are our engineers engaging with AI coding agents in each session, and what does session depth tell us about proficiency versus unproductive looping?
What it shows
Session Depth tracks the average number of prompts or turns per AI coding agent session, measuring how deeply engineers engage with the agent to complete a task. A session with 1-2 turns is a single-prompt exchange. A session with 15-20 turns represents an extended back-and-forth where the engineer is actively steering the agent through a complex task.
Why it matters
Session Depth is a proficiency signal. Very shallow sessions (1-3 turns) can mean the engineer is using the agent for simple, single-step tasks, which is legitimate but limits the leverage available. Very deep sessions without corresponding Outcome Success may indicate the engineer is struggling to steer the agent toward a useful result, a training gap rather than a tool gap. When read alongside the Agent Effectiveness Score and Outcome Success rate, Session Depth helps engineering leaders distinguish productive deep engagement from unproductive loops.
The Larridin angle
Session Depth also has a direct cost implication: longer sessions accumulate more tokens, which matters in agentic architectures where context accumulation across turns drives the majority of cost. A high Session Depth alongside low Outcome Success is the signature of expensive, unproductive looping, exactly the pattern Larridin's Token Spend & Insights layer is built to surface.