# Cognition Links Devin Usage Limits to Estimated Engineering Output

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/34146
Published: 2026-10-17T18:20:41.000Z
Updated: 2026-10-17T18:20:41.000Z
Section: Technology

> The estimator compares completed sessions with estimated human engineering hours, while Devin Coach flags habits that can waste usage.

Cognition is using estimated engineering work per session cost to guide which Devin users may qualify for higher usage limits, making completed work a factor in access decisions.

The estimator reviews completed sessions, estimates how many hours a human engineer would need to do the same work without Devin, then divides that estimate by the session cost. Engineers graded “healthy” or “satisfactory” may qualify for higher limits. In this approach, raw token use is a weaker signal: employees using more compute need to show more estimated output to qualify for additional access.

Cognition is also rolling out Devin Coach, which flags habits that can waste usage. Examples include vague prompts, choosing a model that costs more than a task requires and bundling separate tasks that could run in parallel. The estimator measures completed work against cost; Coach focuses on ways users can change how they use the tool.

Calvin Giroud, Cognition’s vice president of customer engineering, said limits should change over time: “The limits that a user should have should not be static. They should be dynamic.”

The approach extends Cognition’s [earlier productivity case for Devin](<https://www.tokenpost.com/news/technology/26537>), which emphasized helping developers take on more work. The estimator builds on a system outlined in June, comparing its time estimates with engineers’ estimates of how long the work would take without Devin.

In an evaluation, the estimates had a 0.74 log-scale correlation across 233 held-out sessions. Those sessions came from a broader group of 258 sessions involving 126 users. The comparison gives the system a way to translate completed sessions into estimated engineering hours and weigh them against their cost.

The result does not mean every session estimate is precise. Individual estimates can commonly be off by two to three times, and the validation relies on engineers’ own time estimates. Estimated engineering hours also do not measure business value: similar amounts of work can have different value depending on what the work accomplishes.



## Links in this article

- [earlier productivity case for Devin](https://www.tokenpost.com/news/technology/26537)
