Skip to main content
Cloud feature. The Autonomy Ladder widgets are available in the AgentMark Dashboard on the Executive Overview.
Whether an organization is becoming more agentic needs a work-level answer, not a tooling-adoption survey. The Autonomy Ladder classifies every merged pull request by how much human oversight its agent sessions actually needed. The classification comes from telemetry your agents already emit, never from self-reports.

​
The levels

Every synced agent session gets a level from three signals stamped at ingest: follow-ups (human turns beyond the initial hand-off: a correction, an answer, or a new ask in the same session all count, because each one means the work wasn’t a single delegated hand-off), denials (tool calls a human rejected at a permission prompt), and run origin (interactive seat vs. cloud or CI worker). A merged PR inherits the minimum level across the sessions that produced it: a follow-up or denial anywhere in the chain means the work wasn’t delegated end to end.

​
Deliberate choices

Approved permission prompts don’t demote. A session that stopped to ask, and got a yes, still counts as Delegated. An approval unlocks a door without changing the agent’s course; only a denial alters it. Penalizing prompts would push teams to turn off permission guardrails to look autonomous, which is the opposite of what the metric should reward. How often sessions run prompt-free has its own tile: Auto-Approved Sessions. Merging is the cohort filter, not a level requirement. The level describes how the work ran; whether it held up is a separate question. The dashboard pairs the ladder with the Clean Job Rate guardrail instead of folding quality into the level. Rising delegation is only good news if quality holds, and a composite number would hide exactly that trade-off. Unclassifiable work stays excluded, never guessed. Sessions synced before the steering signals existed have no honest classification: an interactive session always records its initial ask as a user turn, so one without any user turns can only be a legacy row. Merged PRs whose sessions all lack a classification, or that have no synced session at all, appear in a separate count instead of silently landing in a level. When coverage is low, the ladder describes the observed slice of your shipped work and says so.

​
Reading the widgets

  • Shipped Work by Autonomy Level: daily merged-PR counts stacked by level. Watch the Delegated and Autonomous bands grow against the Assisted and Supervised base.
  • Delegated+ Share of Merged PRs: the single trend number. Of classified merged work, the share at L3 or L4. Read it beside Clean Job Rate (merged, not reverted, zero follow-ups). Delegation climbing while clean-job holds means warranted trust; delegation climbing while clean-job slips means rubber-stamping.

​
Signals reference

All signals aggregate work, never rank people. The dashboard’s no-per-developer-data rule applies to the ladder like everything else.

​
Have questions?

Reach out any time: