Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

Devin now runs on a budget — and it can cut you off mid-task

On November 2, Cognition gave Devin hierarchical budgets: one company-wide ACU budget splits down to departments, projects and teams, up to three levels deep. When a budget runs out, the running session stops — warnings and in-product top-up requests are the cushion, and efficiency-based auto-approval rewards engineers who spend well. A soft-limit fallback to cheaper models is still in development.

A business team gathered around a conference table reviewing printed charts and financial documents

Spending first: Devin finally budgets the way companies are organized

On November 2, 2026, Cognition added hierarchical budgets to Devin, for Devin Enterprise customers. The same day, Cognition announced the launch on LinkedIn, and Lukas Burger — who works across go-to-market and engineering at Cognition — described the feature to The New Stack in an interview. This update adds no new model and no new capability. It changes how the money is managed: one company-wide ACU (Agent Compute Unit) budget can be split down to departments, projects, and teams, up to three levels deep.

Why now? Burger's answer in the interview was direct: the feature came out of work by Cognition's forward-deployed engineers — the people who sit on-site with customers and help enterprises roll out Devin. The old controls were organization-level limits, per-user tiers, and a separate set of settings for Devin Desktop. That was fine for one team, but it stopped holding up once Devin spread across a whole company. Burger's own words: "That's fine for one team. It doesn't hold up when a bank is dividing a large ACU commitment across hundreds of groups." Some customers had even built their own internal tools to parcel out capacity. Cognition says it designed the feature with some of the world's largest enterprises as design partners, including customers managing millions of ACUs — though Burger declined to give individual customer spending figures.

That background is worth keeping in mind: hierarchical budgets were not dreamed up by Cognition as a governance feature. They arrived because enterprise customers pushed Devin from pilot to company-wide use, and the bill broke first. When an agent goes from "a few people trying it" to "hundreds of people using it every day," the first problem to surface is not a lack of capability — it is money without structure. That is the starting point for this article's judgment: the significance of hierarchical budgets is not in how clever the feature is, but in what it marks — AI coding agents formally entering the era of "managing money by organization."

How the three-level tree grows: one root, three levels, every parent in charge

The official documentation lays out the structure plainly. Each enterprise account has one budget tree. The root is the Account budget, child budgets sit under it, and the whole tree can be up to three levels deep. A budget can hold one organization or several, but each organization belongs to exactly one budget; new organizations go under the Account budget. A budget's limit is a monthly ACU cap for one monthly window; a null limit means unlimited.

The tree runs on four hard rules, and understanding them is understanding hierarchical budgets. First, usage rolls upward: a child budget's usage also counts toward its parent and the Account budget. Second, an empty parent stops the whole branch: when any budget on the path runs out, usage in that branch stops — not just the child budget, the entire branch below it. Third, child limits can add up to more than the parent limit: the sum of several child budgets may exceed the parent's limit, but the parent still caps their combined usage. Fourth, a budget with no limit is not free: it is still capped by its parents.

The management entry point is Enterprise Settings, under Budgets. Enterprise admins manage the whole tree, including the Account budget; they can also share a budget with a person or group, making them delegated admins for it. A delegated admin can create budgets under it and set limits and temporary limits on those child budgets, but cannot raise the shared budget's own limit — only whoever manages the parent budget can change that. Delegation requires separate enablement; customers need to reach out to their account team. In terms of power, hierarchical budgets mirror the org chart itself: the center sets the total, departments handle allocation, power shrinks as you go down, but every level has its own owner.

One more point: local and cloud usage count against the same limit. Local usage on Devin Desktop and Devin CLI and cloud usage in Devin sessions all count toward one limit, and once a limit is reached, new work is blocked on all surfaces. Devin Review usage counts toward budgets but does not count toward user limits. Budgets and user limits are enforced independently — new work is blocked if either limit is reached.

Two official examples, with the math done

The documentation gives two examples, and all the numbers are from its Acme Corp sample — useful precisely because they let us do the math. The first is delegated departments. Acme has 10 departments, and central IT sets each department's total: the Account budget is 100,000 ACU, Engineering gets 30,000 ACU, Marketing gets 5,000 ACU. Bob, the Engineering department admin, can set limits for the Payments and Search teams under Engineering — for example, a temporary 6,000 ACU limit for Payments during a busy week — but he cannot raise Engineering's own 30,000 ACU limit. Only Alice, the central admin, can do that. Usage in every Engineering org counts toward the 30,000 ACU limit, even when the child budgets add up to more.

Devin hierarchical budget tree (Acme Corp example from the official docs): Account budget 100,000 ACU, split into Engineering 30,000 and Marketing 5,000, plus a Migration initiative pool of 5,000 ACU

The second example is a project pool. Every developer ordinarily gets 200 ACUs a month. For a migration initiative, Acme sets aside 5,000 ACUs and lets people on the project use more: the Developer tools orgs sit in a Developer tools budget with a 20,000 ACU limit and a 200 ACU default tier, while the Migration initiative budget has a 5,000 ACU limit and a 1,000 ACU default tier. Carol, a developer on the project, can use up to 1,000 ACUs in the Migration org and up to 200 ACUs in the Developer tools orgs — and the whole project can never use more than 5,000 ACUs in a month.

Put together, the two examples show the two uses of hierarchical budgets: dividing money by org chart, and ring-fencing money for a project. The first answers "who spent how much"; the second answers "don't let daily spending crowd out the priority work." The logic will feel familiar to anyone who has worked in corporate budgeting — central totals, delegated authority, dedicated funds — except the thing being managed is agent compute, not travel expenses. One caution: every number here is a sample from the docs, not pricing and not a recommendation. Cognition has not disclosed pricing for the feature, nor how many customers use it.

When the budget runs out, the session stops: the hardest rule of all

The rule developers will care about most is this: when a budget runs out, the running session stops. The documentation's wording is: when a budget runs out, usage stops for its orgs. New sessions cannot start, and running sessions sleep; users see "Increase your organization's usage limit." Requests from the Devin CLI and Desktop local agent are blocked, with the message "Your organization has reached its monthly usage limit. Ask an account admin to raise it, or wait for the limit to reset next month." New Devin Review reviews are blocked too.

Note the word "sleep": the session is not deleted, it is parked in place, and it can resume when the next monthly window starts, or when an admin raises the limit or sets a temporary limit. The monthly window is not the calendar month — it follows the contract's billing cycle. If the billing phase starts on the 15th, the windows run from October 15 to November 14, then November 15 to December 14. A temporary limit replaces the budget limit for the rest of the current window only; the permanent limit returns in the next window. A temporary per-user override likewise expires at the end of the current window.

The other side of this rule is that it turns "saving money" from advice into enforcement. Before, going over budget meant, at worst, a large bill at the end of the month. Now going over budget means the work in your hands stops on the spot. For developers, that means the work schedule has to follow the budget window: plenty of allowance early in the window, thin allowance at the end; important work should not be scheduled for the last days. For admins, it means a limit is not "set and forget" — set it too tight and the whole team stops together at month's end; set it too loose and hierarchical budgets become decoration. Hierarchical budgets turn a billing problem into a scheduling problem, and that is their most underestimated effect.

Advance warnings and in-product requests: don't let the stop come as a surprise

So that "the session suddenly stopped" does not become a daily event, Devin provides two layers of cushion. The first is warnings: engineers get a notification when they reach around two-thirds of their limit, then persistent banners as they approach and reach it. The second is requesting more: engineers can request more ACUs without leaving the product — no email, no approval form. The request entry point is the "Request more" button on the ACU consumption chart in the My analytics page, or directly from the limit warnings as they approach or hit their limit. They pick an amount, optionally add a reason, and can track or cancel the pending request. Every approved increase is temporary, applying for the current billing cycle only.

The approval rules are set by each tier's administrator, in three flavors: manual approval, where every request waits for admin review; always approve, where requests are auto-approved up to the tier's maximum auto-approve allocation; and approve based on efficiency, where requests are auto-approved up to the ceiling, but only while the member's Efficiency Score is Healthy or Satisfactory, and only in small increases ahead of their current usage. Auto-approval only ever works in the member's favor: it never denies automatically — denying is always an admin decision, and a request that cannot be auto-approved simply waits for review. A request above the auto-approve ceiling is granted up to the ceiling so the member can keep working, and the remainder stays pending. A permanent override set by an admin is never replaced by an auto-approval.

One design detail deserves credit: requesting more is a first-class feature inside the product, not an afterthought in the admin console. Engineers do not have to go find someone after their session stops, and admins do not have to process a pile of requests at month's end. But the cushion is not insurance — if admins sit on requests for days, or the auto-approve ceiling is set too low, the month-end stop still arrives. Warnings and requests buy time; the time has to be used to make decisions, not to postpone them.

The Efficiency Score link: people who spend well get more automatically

Hierarchical budgets did not ship alone — they build on the Efficiency Score, introduced in October. The score estimates how many productive engineering hours each Devin session produced relative to the ACUs it consumed: how much real engineering output the compute bought. Contemporary reporting puts the launch at October 19, 2026, with a 0–100 scale; the official docs present the result as four grades — Healthy, Satisfactory, Needs improvement, and Unknown (for members whose recent usage is mostly local or too small to grade). This article cites the 0–100 figure and the October 19 date from contemporary reporting, while the docs themselves only show the four grades — readers can treat the score as the reported scale and the grades as what the product actually displays.

The link works through "approve based on efficiency." An admin can switch on this policy for a tier: when a member's Efficiency Score is Healthy or Satisfactory, limit-increase requests are auto-approved up to a ceiling the admin sets — and only in small increases ahead of their current usage, not a lump sum. Members in good standing near their cap also get increases granted automatically, without manually filing a request. Members can see their own score — along with tips for improving it — on their My analytics page, and Devin Coach helps them improve it as they work.

The thinking is clear: budgets should not cut everyone equally; money should follow output. People who spend well get more; people who spend poorly get held back until they improve. But the controversy is just as clear: the Efficiency Score estimates "productive engineering hours." It does not measure code quality, and it does not measure business outcomes — The New Stack's reporting states that explicitly. A session that produces many "productive hours" has not necessarily produced good code; an engineer with a low score is not necessarily slacking — they may be working on the hardest tasks, such as legacy systems, research, or documentation, where output is slow to show. Tying a score directly to allowance turns one estimate into a basis for allocating resources, and the estimate's errors become differences in treatment.

One more detail is worth noting: the score looks at recent cloud-session usage, so early mistakes do not count forever. Cognition has said it validated the hours estimates against engineers' own assessments, and that the model tracks well in aggregate. But validation is one thing and turning a score into allowance is another. My advice: before switching on "approve based on efficiency," admins should look at a month of score distribution and check whether the low scorers are genuinely inefficient or simply holding the hardest tasks. Engineers should know that when their score is low, the product offers improvement tips first — use that path before complaining about the allowance.

Still in development: the soft limit that downgrades instead of stopping

The biggest missing piece of hierarchical budgets is one Burger admitted himself in the interview: the option to fall back to lower-cost models after a soft limit, "so important work can finish," is still in development, with no date given. In other words, today's budgets have only one ending — a hard stop. Either the allowance covers the work and it continues, or the allowance runs out and the session sleeps. The middle option — "allowance nearly gone, switch to a cheaper model and finish the job" — does not exist yet.

That gap matters most under the "session stops mid-task" rule. A hard stop hurts not because of the money, but because of the work in progress: a migration half-run, a refactor half-done. The cost of stopping midway is not just the compute to re-run it, but the engineer's time to get back into context. Warnings and requests soften the blow, but they do not fix it — warnings only tell you in advance, and requests still wait for an admin to approve. The soft-limit downgrade is seen by many as the most important missing piece precisely because it turns "stop" into "slow": the work still finishes, just on a cheaper model at a slower pace.

Until the soft limit ships, enterprises have only two levers. First, schedule critical work early in the window, not when the allowance is nearly gone at month's end. Second, get the temporary-limit approval flow running smoothly, so "request, then approve" takes minutes rather than dragging into the next day. Burger said the point of the soft limit is that "important work can finish." Before it ships, that responsibility sits with the admin's process, not with the product's features.

Verdict: a tool for managing money — and a mirror of how people use it

The real question about hierarchical budgets is not how they divide money, but how they will change the way people use Devin. Once budgets have levels and owners, usage is no longer "the company's bill" — it is "a department's, a project's, a person's bill." Every session is tied to a specific budget, and companies can attribute spend to a team, project, or business unit and map it to internal billing codes or departmental chargeback. For finance, that is good news: the agent's bill used to be one blurry number; now it can say who spent each unit and for which project.

For engineers, what the mirror shows may be less comfortable. Your usage gets tiered, scored, and compared with colleagues'; your session can stop because someone else's budget ran out first — remember, when a parent budget runs out, the whole branch stops, even if your own budget still has room. Hierarchical budgets turn "personal efficiency" into data the organization can see. That is progress for management and pressure for the experience. Cognition's design leaves cushions — warnings, requests, efficiency-based auto-approval — but those cushions only work if admins actually manage, rather than setting a limit and walking away.

Three suggestions for enterprise admins. First, do not set the limits tight on day one: run the first month with generous limits, watch the real usage distribution, then tighten by data. The docs offer a "set up with recommendations" flow that analyzes the last three billing cycles of per-member usage, suggests a complete tier configuration, and simulates who would be over their limit and the projected request volume — use it for the first version instead of guessing. Second, push temporary-limit approval down to department heads; if every request piles up with the central admin, the hierarchy is decoration. Third, give critical work its own budget, separate from daily development — the migration-initiative example is the template: dedicated funds mean priority work does not get crowded out by everyday spending.

For developers, the advice is simpler: treat the budget window as part of scheduling. Early in the window, allowance is plentiful — a good time to start long tasks. Late in the window, allowance is thin — a good time for light work. When a limit warning appears, request more inside the product right away and add a reason. If your Efficiency Score is Healthy or Satisfactory, auto-approval will help, but do not treat it as unlimited — auto-approval has a ceiling, and anything above it still waits for a human. And most importantly: do not start a migration that takes hours on the last day of the window. Stopping halfway costs far more than starting a day later.

What it means for vibe coding developers

One honest note first: hierarchical budgets are for Devin Enterprise customers only. Individual and indie developers cannot use them now, and will not feel their effects directly any time soon. But that does not make them irrelevant to vibe coding — they signal where the whole industry is heading. When agents move from personal tools to organizational productivity, spending management shifts from "look at the bill afterwards" to "divide the budget beforehand." Devin is doing hierarchical budgets today; other agent platforms will follow with similar controls in different forms. The direction is the same.

For indie developers, what can be copied right away is the method, not the feature. Set yourself a monthly ACU or token budget, split by project: one pool for everyday small tasks, one for the priority project, one for trying new tools. Pools do not borrow from each other, and the priority project's pool is kept full on its own. This needs no enterprise feature — a spreadsheet does it — but it solves the same problem: do not let everyday spending quietly eat the budget of the work that matters. Many indie developers only discover at month's end that the money went on "quick questions here and there," while the important project ran out of allowance. The hierarchical-budget mindset fills that hole in advance.

For developers using Devin inside a company, the first thing to do after budgets go live is to open the My analytics page and look at your usage, your limit, and your Efficiency Score — which budget you sit under, how much is left, and which grade you hold. Nobody may have told you these things before; now the product shows them all. The second thing is to ask your team lead three questions: who approves temporary limits for our budget? Is auto-approval switched on? What is the ceiling for efficiency-based approval? Getting those three answers is worth more than ten explainers. Budget rules are not frightening when they are strict; they are frightening when they are opaque — and knowing the rules is what lets you plan your work.

Finally, back to the industry picture. Devin's hierarchical budgets and Claude Code's Code Mode are two seemingly unrelated things from the same week that point in the same direction: one manages "how much the agent spends," the other changes "how the agent works." One governs money, one governs work — together they are the two lessons agents must learn to graduate from "fun tools" to "serious productivity." A productivity tool is never judged only by how capable it is; it is also judged by two things: whether its spending has structure, and whether its work has rules. Devin and Claude Code are each teaching one of those lessons in the same week.

Views 4Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
A developer types at a dual-monitor workstation; the left screen shows a dark code editor, the right a testing dashboard with green checkmarks and bar charts
News
Did Your Agent Get Better or Worse? AWS Open-Sources UAEF to Measure Agents from Every Framework with One Ruler

On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.

Developer WorkflowProduct LaunchTesting & Quality
Against a black background, a conductor seen from behind raises a baton, with swirling trails of light circling beneath him like an orchestra being conducted
News
One API for 45 Models, and It Switches Providers When One Fails: Perplexity's Orchestrator Wants to Be the Switchboard

On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.

Developer WorkflowProduct LaunchModel Updates