Devin Coach and Efficiency Scores: Spend Limits That Follow Usage

I have not used Efficiency Score, Devin Coach or efficiency-based approvals myself. I am on the Devin Max plan, and outside /scan I rarely run into usage limits. A single /scan run once used up to roughly 60% of my weekly limit. I find that normal: /scan sends dozens of agents across the whole repository in one go, purely to find bugs, where saving tokens matters little.

On October 19, 2026, Cognition introduced three enterprise features for Devin that tie spending limits to how well each engineer uses the agent: Efficiency Score, Devin Coach and efficiency-based approvals. What follows is drawn from Cognition's announcement rather than from a measured test, and my own observations concern everyday Devin usage, not the new features.

What Cognition shipped

Efficiency Score is a 0 to 100 rating of an engineer's recent Devin usage. Engineers see it in My Analytics, under Session efficiency, as one of four grades: Healthy, Satisfactory, Needs Improvement, or Unknown when there is not enough data. Supporting metrics include merged PRs per day, ACUs per merged PR and sessions with delivered outcomes.

Devin Coach reviews a prompt before it runs. It flags a vague request, an expensive mode for a simple task, unrelated tasks in one session or non-work use, and suggests a fix such as making a plan first or switching away from Ultra. Cognition says Coach never rewrites a prompt silently; admins choose inline suggestions, a required confirmation, or no checks.

Efficiency-based approvals let an admin select “Approve based on efficiency” in Usage policies and set a ceiling. When an engineer with a Healthy or Satisfactory grade nears their limit, Devin raises it automatically up to that ceiling. Other requests go to an admin, and Devin never denies one automatically. Cognition says auto-approvals are available today on the Devin Enterprise plan.

Why the score matters more than the cap

The idea behind the score is that merged pull requests undercount useful work. For each session, Devin checks whether useful work was produced: a merged PR or, for sessions without one, a classifier's judgment that real progress was made on triage, analysis or an investigation. It then estimates how many hours a human engineer would have needed for the same work.

Those productive hours become two ratings. Efficiency is productive hours per ACU, the Agent Compute Unit Cognition uses to combine model and infrastructure cost. Volume is productive hours per day. Efficiency carries more weight: volume can add at most a fifth of the points left between the efficiency rating and 100.

I mostly use the Fusion model, and I like its output a lot. It feels like picking Opus in the model picker, yet the drop in my weekly usage has been small compared with how difficult the tasks were: even large refactors in cloud tasks spanning two repositories moved my weekly usage by about 4–5% at most once they were done. That is my impression, not a measured comparison.

What the announcement leaves open

The score depends on Devin's own estimates. Cognition says it validated the hour estimates against engineers' assessments of their sessions, but that dataset is not public, and the classifier that credits sessions without a PR is not described in detail. The example of an engineer going from 30 to nearly 80 merged PRs a week is a vendor-reported case, not a controlled comparison.

The features are built for organizations with admins and usage policies. Teams on smaller plans may not see auto-approvals at all, and Cognition says budgeting controls that let department leaders distribute limits to sub-teams are still to come.

Where I would draw the line

For my individual usage, the bigger cost driver appears to be /scan-style fan-out rather than everyday work. That is an inference from my own use, not a test of the new features. Since Efficiency Score, Devin Coach and efficiency-based approvals are Enterprise-only, they look aimed more at Enterprise teams whose admins set usage caps than at an individual on a plan like mine.

A useful next test would record the plan, the Efficiency Score grade over a few weeks, which Coach suggestions appeared and were accepted, and whether ACUs per merged PR changed after following them.