Signals & Insights
Intent Fidelity
Insights · WorkspaceWhether completed work matches what was actually asked for — using the agent's own account of its work at handoff, not an automated computation. When a strategic bucket's average drift creeps up, the next agent claiming a same-bucket objective sees a stay-scoped advisory before writing any code.
Avg drift score
A 0–100 self-assessment the agent submits alongside every acceptance request — 0 means perfect alignment with the stated intent, 100 means complete deviation. Nothing computes this in the background; it's the agent's own read on its work, taken at the moment it has the richest context, averaged over the trailing 30 completed objectives workspace-wide. Shown as a color-coded badge (green under 20, amber 20–40, red above 40) next to each objective in the Recently Completed list on the Insights Deep Dive tab, and as an average on each agent's card in the Team tab once that agent has at least 5 scored objectives.
Healthy
Average is below 20. Agents consistently report clean alignment with what was asked, and they're reporting it with full visibility into their own diff and acceptance criteria at handoff.
Watch for
Average is 20–40 — worth monitoring, most often on broadly-scoped objectives. Above 40, treat it as active: agents are regularly telling you their delivery diverged from the ask.
Scope drift rate
The percentage of objectives completed in the trailing 90 days whose most recent rejection was categorized wrong_scope or partial_implementation. Unlike drift score, this needs no self-report — it reads directly off rejection history.
Healthy
Rate is under 15%. The scope written into an objective is holding up through execution and review.
Watch for
Rate above 15% is worth a look; above 30%, treat it as urgent regardless of what the drift score says — a large share of completions are landing outside what was originally asked for.
Spec stability
The percentage of completed objectives where the acceptance criteria at done time match the snapshot captured when the agent claimed the objective. A mismatch means the spec moved after an agent was already executing against it.
Healthy
Stability is 80% or higher. Objectives are locked before agents start work, so mid-run rewrites aren't forcing rework.
Watch for
Stability is 60–79% — worth watching; below 60%, treat it as active. Acceptance criteria are being edited after claim, one of the more common causes of a high drift score and a lower first-pass rate, since the target moved after the agent had already started.