Cognition Adds Hierarchical Spending Controls to Its Devin AI Coding Agent

When an AI coding agent can consume resources across long, multi-step tasks, a company-wide usage total tells finance teams very little about who is spending or what the work is producing. Cognition’s new hierarchical budgets for Devin bring those decisions into the product, giving enterprises controls for allocating Agent Compute Units (ACUs) across departments, teams, projects and individual engineers.
The practical trade-off is clear: organizations gain more visibility and control, while developers risk having a Devin session stop if its budget runs out mid-task. Cognition has added alerts and approval requests to reduce that chance, but the spending limit is enforced at the session level.
Budgets built for deployments beyond one team
Earlier controls such as organization limits, per-user tiers and separate Devin Desktop settings could work during a pilot. They became harder to manage as deployments spread across a company. Cognition’s Lukas Burger, who works across go-to-market and engineering, says the budgeting feature came from customer discussions led in part by the company’s forward deployed engineers, who work directly with customers.
For a large enterprise, the challenge is distributing a substantial ACU commitment among many groups. Burger describes banks dividing large commitments across hundreds of teams, with some customers building internal tools to allocate capacity among business lines. In his view, the shift from pilot to broad use changes the ownership model: Devin no longer belongs to a single engineering group, and multiple departments need their own spending authority.
Cognition developed Budgets with some of the world’s largest enterprises as design partners. The customers involved were managing millions of ACUs, the consumption-based units used to measure Devin’s work. Burger hasn’t disclosed individual customer spending figures.
Allocation follows teams and projects
An administrator can create an overall budget and divide it among business units. Leaders in those units can then distribute their allocations to teams and projects, with the hierarchy extending up to three levels. Central IT retains visibility across the structure and can set limits, while budget owners handle decisions closer to the work.
Projects can receive their own allocations even when contributors come from different departments. That makes it possible, for example, to track a legacy modernization effort as a distinct cost center instead of splitting its consumption among the engineers’ home teams. Companies can also apply separate controls to Devin’s cloud environment and its use on engineers’ desktops.
Each Devin session is associated with a budget. That gives managers a way to examine consumption at the level of a project, team or business unit rather than relying only on a company-wide total. Some customers connect those costs to internal billing codes or charge them back to departments.
Requests and limits shape the developer experience
Devin alerts engineers on their homepage when they’re approaching a budget limit. If they need more capacity, they can request additional ACUs within the product. Budget owners choose whether those requests require manual review or can be approved automatically.
The sharp edge remains the hard stop: once a budget is exhausted, the active session ends. Cognition is working on an option that would let teams switch to lower-cost models after reaching a soft limit, allowing important work to continue. Burger says that capability is still in development, so advance warnings and budget increases are the available ways to avoid an interruption.
That design puts pressure on administrators to set limits with enough room for productive work. Cognition says it wants customers to control spending without setting caps so low that engineers are routinely blocked. The balance will depend on how accurately teams can forecast usage and how quickly budget owners respond when work exceeds expectations.
Efficiency scores can trigger more capacity
Administrators can enable efficiency-based approvals, which raise an engineer’s limit automatically when their Efficiency Score is high, subject to a ceiling set by the administrator. Cognition introduced the scoring system in October. It estimates productive engineering hours relative to Devin consumption, giving budget owners a way to allocate more capacity to engineers who appear to get more work from their usage.
That score needs careful interpretation. It estimates efficiency; it doesn’t directly measure code quality or business outcomes. Companies will need to pair it with other indicators before treating a high score as proof that additional consumption is paying off.
For that reason, the controls matter most when they connect usage to results. Burger says customers commonly track merged pull requests, productive engineering hours and DORA metrics. Teams that get the most value, he says, also look at how quickly roadmaps progress, whether unplanned work declines and how much engineering toil is reduced.
Usage visibility is only half the job
Consumption pricing can expose waste, but it can also make bills harder to predict when agents tackle complex tasks. Burger says one source of surprise is simply not knowing where usage accumulates. Lower token prices don’t automatically mean a smaller bill if the agent runs through many steps or handles more work.
A round of price cuts doesn’t necessarily translate into a smaller bill, which makes attribution and measurement central to enterprise adoption. Cognition’s Budgets put those controls inside Devin, letting organizations assign spending limits and see which initiatives consume ACUs. The product can show where money went; customers still have to decide whether the output justified the cost.
Burger says customers who can’t see value in a consumption-priced product stop using it, which is why Cognition has focused on cost controls and measurement. That puts the company’s feature in a broader operational shift: teams are building financial practices around AI agent usage, much as they did for cloud infrastructure.
AI spending becomes a finance responsibility
Burger says agent budgeting is already developing into a discipline comparable to FinOps. Finance teams are asking familiar questions: who is spending, where the money is going and whether it produces measurable value. He expects the transition to happen faster than cloud FinOps did because many companies already have cost management processes they can adapt.
The FinOps Foundation’s 2026 “State of FinOps” report found that 98% of 1,192 respondents manage AI spending, up from 31% two years earlier. Connecting that spending to returns remains a work in progress. Devin’s approach embeds allocation and tracking in the agent product, while the business still has to define what counts as a worthwhile result.
Burger expects strategic finance and financial planning and analysis teams to take on a larger role as AI budgets grow. Their task will be to connect consumption to impact on the profit and loss statement. For engineering leaders, the immediate takeaway is more operational: set budgets around real ownership, watch usage against delivery measures, and leave enough headroom that a productive session doesn’t end before its work is done.
The post Your AI coding agent now runs on a budget — and it can cut you off mid-task appeared first on The New Stack.