Cognition adds hierarchical budgets to Devin for enterprise AI spend

Cognition added hierarchical budgets to Devin on 2 November. An enterprise admin now sets one account limit at the top. Child budgets sit under it, up to three levels deep. Group budgets, project budgets, and per-user limits live inside them.

Each budget holds a pool of Agent Compute Units, the measure Cognition uses for Devin's work. A budget covers one monthly window. An admin can also give a child budget a temporary bump that lasts to the end of the current window.

How the tree works

The root is the account. One department head might hold the Engineering budget below it. Smaller orgs sit inside that one. The department head can set limits for every org under them. They cannot see or touch another department's budget.

A busy launch team can get a temporary allocation from their parent. Cognition's blog gives the example of a Payments team at 10,000 ACUs. The parent grants a 5,000-ACU bump for one window, taking them to 15,000. Sibling teams stay bound by the combined parent limit.

Project budgets work the same way. Cognition describes a security hardening sprint that pulls from a dedicated 5,000-ACU pool. A developer can also hold a smaller discretionary budget beside it. They cannot spend from the security pool unless the work belongs to that project.

When a budget empties

An engineer sees a warning on their homepage as they near the limit. They can file a request for more ACUs without leaving Devin. Each budget's admins choose whether to clear those requests by hand or in advance.

If the budget runs out while Devin is working, the session stops. The New Stack reports that Cognition is building a soft-limit option. It would switch to a lower-cost model after a cap, so a long task can finish. That option is not out yet.

Cognition's Lukas Burger told The New Stack where the feature came from. It grew out of forward deployed engineers working on site with banks and other enterprises. The old org limits and per-user tiers broke down once one commitment had to cover hundreds of groups.

Where the cost lands

For a team paying for Devin, a per-group budget answers the question of where spend went. Each session is tied to one budget. Finance can see the Payments team's bill against the Analytics team's bill, instead of one company-wide total.

For an engineer on a tight budget, the cost is new. A session can halt in the middle of a task. The alerts and in-product requests are the only way to avoid it before the soft-limit model-switch option arrives.

Agent budgets look a lot like what cloud bills got in 2018. The New Stack cites the FinOps Foundation's 2026 State of FinOps report. It finds 98 percent of 1,192 respondents now manage AI spend. That is up from 31 percent two years ago.

Source

Cognition on the hierarchical budgets feature. The New Stack on the design partners and the FinOps context.