Thesis: Cost controls for coding agents are moving from uniform caps toward per-user allocation tied to estimated productive work.
Observed facts: Cognition introduced Efficiency Score, Devin Coach and efficiency-based approvals for enterprise users. Observed facts: The 0–100 score uses estimated productive engineering hours per Agent Compute Unit and a smaller volume contribution; qualifying sessions without a PR may count when a classifier judges useful progress. Mechanism: Devin classifies recent work, estimates its human-equivalent engineering hours, relates those hours to compute spend, and uses the resulting grade in an approval rule. Healthy or Satisfactory users nearing their cap can receive an automatic increase up to an admin-set ceiling; other requests remain subject to admin review. Coach separately intervenes before a run with suggestions about prompt scope, mode choice or session structure.
Why now: As teams adopt coding agents, uniform caps can constrain high-value usage while creating review work; Cognition is productizing a way to differentiate access and coach behavior. First implication: Enterprise teams can trade blanket restrictions for differentiated capacity, but they are also making an estimated productivity score part of spend governance. Evidence boundary: These are Cognition’s descriptions of its own features and scoring approach; the announcement does not independently establish that estimated hours or grades reliably measure engineering value, nor does it report comparative results. First watch condition: Look for customer evidence on whether score-based approvals reduce admin review without rewarding noisy or misclassified work.
This starts a distinct analytical record because no prior New Runtime record is supplied for this specific mechanism: per-engineer estimated efficiency grades directly govern automatic Devin limit increases.