Cognition teaches Devin to grant limits based on efficiency: a 0–100 scale

Cognition, the company behind the Devin agent, is urging enterprise customers to drop flat caps on AI spend and allocate budget by return instead. It has released three features for this: Efficiency Score (a 0–100 rating for each engineer), Devin Coach (prompt suggestions before a run) and auto-approval of limit increases for efficient users. Auto-approval is already available to all customers on the Devin Enterprise plan. Cognition likens a blanket limit to a hiring freeze: it constrains the people who get the most out of the agent, and it does not tell everyone else what to do differently.

How Efficiency Score is calculated

Cognition's base unit is the Agent Compute Unit (ACU), which combines LLM cost and cloud infrastructure cost. The score is built from recent sessions rather than an all-time average, so old mistakes do not follow a user forever.

First, Devin determines whether a session produced a useful result. If there was a PR, it must have been merged. If there was no PR, a small classifier decides whether the agent made progress on the task, such as diagnosing a bug or answering an analytical question. That way, investigation and planning also earn credit. Then, for each counted session, Devin estimates how many hours the same work would have taken a human. Cognition checked these estimates against engineers' own estimates for their sessions.

The result combines two metrics:

• E (efficiency): productive hours per ACU;
• V (volume): productive hours per day, regardless of cost;
• formula: Score = 100 × (E + λV(1 − E)), where λ = 0.2;
• both ratings are normalized to a 0–1 scale and flattened at high values;
• grades: Healthy, Satisfactory, Needs Improvement, and Unknown when there is too little data.

Examples based on the last 200 ACU show how it works. A "typical" engineer (E = 0.50, V = 0.50) scores 55. A power user with the same E but V = 0.95 scores 60. A "heavy but wasteful" user (E = 0.27, V = 0.95) lands at 41. A careful engineer with low volume (E = 0.75, V = 0.17) scores 76. Volume adds no more than a fifth of the points remaining up to 100, so volume alone cannot earn a high score.

Coach and auto-approval

Devin Coach reviews a prompt before it runs. Cognition identified typical beginner mistakes: a request that is too broad, the most expensive mode used by default, and several unrelated tasks in one session. For example, for "rewrite billing for usage-based pricing," Coach will suggest drawing up a plan first. For a button text change, it will recommend the regular mode instead of Ultra. For unrelated requests, it will suggest a new session or child sessions, and for clearly personal requests such as vacation planning, it will remind the user that this is not work usage.

Under the hood, a small LLM classifier works with a fixed set of patterns. It was tuned on several hundred labeled examples, with an emphasis on cases where a good prompt should be left alone. The admin chooses the level of intervention: suggestions that can be ignored (the default), mandatory confirmation for a prompt flagged as a new session, or turning Coach off entirely. Coach does not silently rewrite prompts. Admin analytics show which suggestions users see, accept and dismiss.

Auto-approval is enabled in Usage policies with the "Approve based on efficiency" option. When an engineer rated Healthy or Satisfactory approaches the limit, Devin raises it automatically, but not above the ceiling set by the admin, and organization-wide limits stay in force. If a request exceeds the ceiling, Devin grants the maximum and passes the remainder to the admin. There are no automatic denials: requests from Needs Improvement and Unknown users go to manual review, where the admin sees the reason, efficiency, model mix and consumption.

Cognition cites two real cases from different companies. In the first, which used manual approvals, an engineer was blocked for four days, and their merged PRs that week fell to a third of the usual level. In the second, which used auto-approval, an engineer hit the limit 22 times, Devin raised it each time, and the weekly count of merged PRs rose from 30 to almost 80. Two observations do not replace a study, and the "productive hours" metric remains Cognition's own internal estimate.

What teams should check before rolling it out

First, look at what your current cap is based on. If it is a single number for everyone, you are most likely penalizing your most useful users. But don't take Cognition's score on faith either: check it against your own reality, for example the share of PRs that reach merge and the ACU cost per merged PR.

It makes sense to start the auto-approval ceiling at a conservative value and expand it based on data. Enable Coach in suggestion mode and watch the share of accepted recommendations: a low share means either a noisy classifier or a training problem. Also note that Needs Improvement goes to a manual queue, so agree in advance who talks to those engineers and how.

In the coming months, Cognition promises flexible budgets: an admin distributes money among department heads, who split limits between teams and approve requests themselves.