You probably use AI for coding like this: open Cursor or Claude Code, describe what you want, the agent generates code, you watch it, point out mistakes and have it fix them, and wrap up a task in minutes or half an hour.
But another way of working is emerging. You hand an entire task to an agent — “migrate this module from the old framework to the new one,” or “figure out why this bug shows up intermittently and fix it” — then close your laptop, go to a meeting, go to sleep. Hours or a couple of days later, you come back and check the results.
This way of working is still early. Goldman Sachs started piloting Devin in July 2025 and is still expanding its use today (CNBC). On independent benchmarks, these agents’ success rates on complex tasks are still far below interactive modes. But the direction is clear: all three frontier labs — OpenAI, Google, and Anthropic — are building dedicated cloud runtime environments for this kind of usage. In a single week in June 2026, three acquisitions landed: SpaceX bought Cursor for $60 billion (Reuters), OpenAI acquired Ona, and Salesforce bought Fin for $3.6 billion. All three point in the same direction.
This article is about why.
Supervised editing and full delegation may look like fast and slow versions of the same activity. They are actually two fundamentally different kinds of work.
With supervised editing, the agent is your tool. You are present the entire time; it accelerates you. The tasks you can do are bounded by your attention and the time you can stay present: fix a function, patch a small bug, add a feature.
With full delegation, the agent is your executor. You give a goal; it explores, experiments, and completes it on its own. The tasks it enables are ones supervised editing simply cannot touch.
Think: refactoring dependencies across ten repos simultaneously, spending two days tracking down a cross-system intermittent bug, migrating an entire module from an old architecture to a new one. These tasks are impossible with supervised editing. They require the agent to explore over long periods, continuously experiment, backtrack repeatedly — a human watching alongside can neither understand what it’s doing mid-process nor help.
This is why long-running matters: it opens up a category of tasks that supervised editing cannot even enter. Speed gains are a side effect. What you can do has changed, and that matters far more than how fast you can do it.
With supervised editing, you can only manage one agent at a time. Your attention is tied to its execution process. Even if it does everything right, you still have to watch, because you need to correct it the moment it goes wrong.
With full delegation, that constraint disappears. You can hand off five tasks simultaneously, each agent running on its own branch. You go to a meeting, do design work, spend time with your kids — then come back and review each result.
You go from being the person who does the hands-on work, to being the person who manages a fleet of agents. The productivity ceiling of these two roles is entirely different. One person watching an agent edit code, versus one person managing five agents each running complete tasks — the gap is several times over.
This shift may not look dramatic from the outside, but it is visceral for people actually using these tools. The time you now spend “watching an agent edit code” will largely become time spent “assigning tasks to agents, reviewing results, and making judgments.” The attention you save can go toward managing more tasks in parallel, or toward work agents cannot yet do.
When agents can take on entire tasks, writing code itself gets absorbed by them.
Your value reduces to two layers: deciding what should be done, and judging whether it was done well.
The role of “developer” is being redefined. From “someone who can write code” to “someone who knows what to build and can judge the outcome.”
This shift runs deeper than the previous two. The first two are about work style and efficiency; this one is about what you make a living on. Writing code becomes the agent’s job. Your core assets become taste, judgment, and a sense of direction: knowing what is worth building, and recognizing what is done well and what isn’t.
For people already writing code, this means the high-order parts of their experience are appreciating while the low-order parts are depreciating. System design, problem diagnosis, trade-off judgment are rising; syntax, boilerplate, repetitive implementation are falling. For newcomers, the barrier to entry has changed: you used to establish your value by writing code. Going forward, writing code won’t need much human doing — you’ll have to build judgment through other paths.
What data does a supervised-editing agent accumulate? What you changed, whether you accepted or rejected its suggestions. Interaction fragments.
What does a fully-delegated agent accumulate? Complete task lifecycles: how a complex task was broken into steps, where it got stuck, how it recovered, how it verified correctness. Task-level, complete records.
The training value of the latter is far higher than the former. To train a next-generation agent that can independently complete complex tasks, you need massive amounts of data on how complete tasks are done — data you cannot derive from “how to complete this snippet of code.”
This explains the underlying logic of the three acquisitions. SpaceX spent $60 billion on Cursor — not for the IDE’s user base, but for the process data of how developers make design decisions. OpenAI acquired Ona (announcement) — for the infrastructure to let agents run long-duration tasks safely. Salesforce bought Fin — for complete records of agents independently completing tasks in customer service scenarios.
The three bought entirely different assets; the underlying logic is the same: whoever owns the complete data and runtime environment for long-running agents secures a position in the next generation of competition.
The three acquisitions are the surface signal. Beneath them, three layers are shifting at once.
The mode of working moves from supervised editing to full delegation. The human role moves from executor to manager. The core skill moves from writing code to judging what to build and whether it’s done well.
For people already using AI to write code, the question isn’t “which tool is better,” but “can I hand off entire tasks?” And more importantly: once I do, where does my value lie?
This shift is still early. The proportion of complex tasks that agents can complete reliably is not yet high. But three frontier labs have voted with their engineering investment, and three acquisitions have voted with real money. Once the direction is clear, all that remains is time and iteration.