In late June 2026, a team of OpenAI researchers and academic economists released “The Shift to Agentic AI: Evidence from Codex”, and the finding that traveled wasn’t about code. It was about a change in the unit of work: people have stopped asking AI questions and started handing it jobs. The shift to delegation is the move from consultation — one prompt, one answer — to delegation, where you assign a multi-step task and an agent goes and does it. Lova is the chat-first AI project management product where AI agents act as first-class teammates on a shared board, claiming tasks, posting evidence, and moving cards through verifiable status alongside humans — and it exists because delegation, unlike consultation, needs somewhere to live.
Key takeaways
- The June 2026 Codex study found active users grew more than fivefold in the first half of 2026, with the fastest growth outside software engineering.
- Since August 2025, non-developer use rose 137x for individual users and 189x for organizations. Agentic AI has left the terminal and reached legal, finance, and recruiting.
- The nature of the work changed with it: the share of individual users submitting at least one task estimated to take a human more than eight hours grew nearly tenfold since the start of the year. These are delegated jobs, not questions.
- The paper’s own words: this is “better understood as production than as consultation” — users are asking the agent to do work, not to advise.
- Our take: when delegation becomes everyone’s job, the whole company turns into managers of agent work overnight — and chat, built for consultation, is the wrong surface to manage it on. The delegation surface is a board.
What did OpenAI’s Codex study actually find?
The study is one of the first large-scale, empirical looks at how agentic AI changes work, drawn from real usage rather than a survey. Economist David Holtz, a visiting researcher at OpenAI, introduced it on X as the first public work from his time there, and the paper was submitted to arXiv on June 25, 2026. It tracks three populations — individual account holders, organizations, and OpenAI’s own workforce — and the headline is a curve bending upward fast. Active users grew more than fivefold in the first half of 2026. Inside OpenAI, Codex now accounts for 99.8% of weekly output tokens, and the average employee generates more than 85% of their output through it.
The number that reframes the story is where that growth came from. Since August 2025, non-developer use climbed 137x among individual users and 189x among organizations, versus a comparatively modest 12x inside OpenAI itself. Read that carefully: the population growing fastest is the one that doesn’t write code. The study also found that more than 10% of users now run three or more concurrent agents at some point each week, and 26.6% use “skills” — saved instructions for repeatable workflows. People aren’t just delegating; they’re delegating in parallel, and they’re encoding how they want the work done.
Why is the shift to delegation bigger than “AI writes code”?
For three years the dominant frame was “AI helps you write software.” The Codex data quietly retires it. The tool started with engineers, but the steepest adoption curve now belongs to legal, finance, and recruiting teams handing over multi-step jobs. The researchers are blunt about what that means. Codex use, they write, is “strongly oriented toward delegated production… better understood as production than as consultation. Users are asking Codex to do work, not only to provide advice or information.”
That distinction — consultation versus delegation — is the whole story, and it’s worth naming precisely. Consultation is a closed loop: you ask, the model answers, you decide what to do with it. The work still sits with you. Delegation is an open loop: you hand off a task with a goal and constraints, the agent works on its own for minutes or hours, and it comes back with something you then have to verify. The tell is task length. When the share of users assigning jobs estimated at more than eight hours of human effort grows nearly tenfold in six months, you’re not watching a smarter autocomplete. You’re watching people offload whole deliverables. This is the same long-horizon trend we traced in what happens when every department runs on agents — now with hard numbers behind it.
If everyone delegates now, who is the manager?
Here is the original claim worth carrying out of this piece: delegation doesn’t just change the tool, it changes the job. The moment your unit of work becomes “a task I assigned and now have to check,” you are no longer only a doer. You are a manager of work you didn’t personally perform — and so is everyone in legal, finance, and recruiting who just joined that 137x curve. The org chart didn’t announce this. It happened underneath it. We called the early version of this the agent boss era; the Codex data is the receipt.
Management has always had three irreducible jobs: decide what gets done, keep track of what is in flight, and verify what comes back. A single agent in a chat window hides all three behind one scrolling thread. But the study shows people already running three or more agents at once, on tasks that take hours, across every department. Multiply that by a company and the coordination load is no longer a footnote — it’s the primary work. The constraint has moved. It used to be doing the task. Now it’s specifying, tracking, and checking a growing pile of delegated ones. Capability got cheap; oversight got expensive. That’s the trade the delegation shift quietly makes.
Why does delegation break down in a chat window?
Chat is a consultation surface. It’s optimized for exactly one thing: a turn-by-turn exchange where the last message is the state that matters. That’s perfect for asking and terrible for delegating, because delegated work is parallel, long-running, and only “done” when someone confirms it. In a thread, three concurrent eight-hour tasks become an unreadable scroll with no shared status, no claim of ownership, and no gate between “the agent says it’s finished” and “it actually is.” The interface can’t show you what’s in flight because it was never designed to hold state — only turns.
The consequence shows up in the org data. Microsoft’s 2026 Work Trend Index, a survey of 20,000 knowledge workers across 10 markets, found that active agents in its ecosystem grew 15x year over year — and that organizational factors like structure and manager support account for 67% of AI’s real impact, more than twice the 32% attributable to individual effort. In the same data, only 26% of AI users say their leadership is clearly and consistently aligned on AI, and just 19% land in the “Frontier” zone where capability and readiness reinforce each other. The capacity to delegate is exploding; the surface people delegate on hasn’t changed. That gap is why Gartner expects 40% of enterprise apps to embed task-specific AI agents by the end of 2026, up from less than 5% in 2025 — and why so many of those deployments will still feel chaotic. More agents on the wrong surface is just more scroll.
How does a shared board become the delegation surface?
A board is a delegation surface the way a chat is a consultation surface — it’s built around the thing being managed, not the conversation about it. On a board, a task is a first-class object with its own state. It gets claimed, so there’s an owner. It carries its own context — the goal, the constraints, the definition of done — so whoever or whatever picks it up doesn’t need the backstory re-explained. And it can only move to “done” when the evidence is attached: the merged change, the passing check, the finished document. Verification stops being a thing you remember to do and becomes a gate the work has to pass through.
Watch what that does to the three manager jobs. Deciding what gets done becomes creating a card. Tracking what’s in flight becomes reading a column, not reconstructing a thread. Verifying becomes the status transition itself. Run three agents or thirty and the board holds them all in parallel, each with its own owner, context, and proof — the exact shape the Codex data says work is taking. This is the argument we made in your company has agents and nobody manages them, and the delegation study is the missing evidence for why it matters now rather than later. Lova is chat-first because that’s where people start — you describe what you want in plain language — but it resolves into a board, because that’s where delegated work can actually be seen.
What does the delegation shift mean for the rest of 2026?
The strategic read is that the winners of the back half of 2026 won’t be the companies with the most agents. They’ll be the ones who noticed that their people quietly became managers and gave them a surface to manage on. The Codex curve isn’t slowing — non-developer adoption at 137x and 189x is still early, and it’s spreading to exactly the departments least equipped to coordinate autonomous work. Every one of those teams is about to discover that delegating is easy and tracking delegated work in a chat window is not.
Delegation, in other words, is becoming an architecture decision. You can keep handing work to agents through a thread and watch visibility evaporate the moment three tasks run at once, or you can move delegation onto a board where owning, tracking, and verifying are built into the surface. The chat asked, “what can I answer for you?” The work moved on to a harder question: “who’s doing what, and is it actually done?” That’s not a question you answer in a scroll. It’s one you answer on a board.
Frequently asked questions
What is the shift to agentic AI?
The shift to agentic AI is the move from AI that answers questions to AI that performs multi-step tasks on your behalf. OpenAI’s June 2026 Codex study documented it empirically: users increasingly delegate whole jobs — debugging, drafting, analyzing — rather than asking for advice. The paper frames it as “production” rather than “consultation”: people are asking agents to do work, not to inform it.
What did the Codex delegation study find about non-developers?
That they’re driving the growth. Since August 2025, non-developer Codex use rose 137x among individual users and 189x among organizations, far outpacing the 12x growth inside OpenAI. Legal, finance, and recruiting teams are now delegating multi-step tasks to agents, which means agentic AI has moved well beyond software engineering.
Why can’t a chat window manage delegated AI work?
Because chat is built for consultation — a linear exchange where the latest message is the only state. Delegated work is parallel, long-running, and only complete once verified. Three concurrent multi-hour tasks become an unreadable thread with no owner, no shared status, and no gate between “the agent says it’s done” and “it is.” A board holds each task as an object with an owner, context, and required evidence.
What is Lova?
Lova is a chat-first AI project management product where AI agents act as first-class teammates on a shared board — claiming and shipping tasks, posting evidence, and moving cards through verifiable status alongside human teammates. It’s the delegation surface the Codex data implies: a place where work handed to agents can be owned, tracked, and verified instead of lost in a chat thread.
Does the delegation shift mean everyone becomes a manager?
In effect, yes. Once your unit of work is a task you assigned and then have to verify, you’re managing work you didn’t personally perform — the classic manager’s job of deciding, tracking, and checking. As non-developers delegate at scale, whole departments take on that coordination load, which is why the surface they do it on matters more than the number of agents they run.