Your AI coding agent now runs on a budget — and it can cut you off mid-task

Cognition, an AI coding agent provider, has introduced hierarchical budgets for its Devin tool to help enterprises manage spending on AI coding. Administrators can assign budgets to different teams, track spending, and approve additional capacity based on engineering efficiency. However, if a budget runs out while Devin is working on a task, the session stops. This feature emerged from conversations with customers, upon whom Cognition relies for feedback on its product development.

Devin's older spending controls, such as org limits, per-user tiers, and separate Devin Desktop settings, no longer suffice for large organizations dividing a large ACU commitment across hundreds of groups. Cognition's budgets address this issue by allowing administrators to divide the overall budget among business units, which then distribute funds to their teams and projects. This model works with the way enterprises already manage spending, keeping guardrails and visibility without the need for individual approvals.

When a budget does run out, Devin stops working, prompting engineers to see alerts on their homepage and request more ACUs (Agent Compute Units) from the budget owner. Administrators can then approve or automatically clear these requests based on engineers' efficiency scores. The system aims to keep engineers productive without intruding on their work.

Cognition is working on additional features, such as allowing teams to switch to lower-cost models after a soft limit, to ensure important work can still be completed. However, the company is cautious not to draw controls so tight that they stifle productivity. The introduction of efficiency-based approvals and a scoring system allows administrators to allocate more capacity to engineers who demonstrate productive usage.

This approach aims to prevent customers from stopping their usage due to high costs, a common issue with consumption pricing models.