On July 9, 2026, OpenAI launched ChatGPT Work — a GPT-5.6 agent that takes a goal, gathers context across your apps, and runs for hours to hand back finished spreadsheets, slides, docs, and web apps. It shipped as a direct answer to Anthropic’s Claude Cowork, and the press had a name ready: the AI super app. An AI super app is a single agent that absorbs a whole job — planning, tool use, execution — and returns finished work instead of chat. Lova is the chat-first AI project management product where AI agents act as first-class teammates on a shared board, claiming and shipping tasks alongside humans. The two are easy to confuse and completely different: the super app is one agent doing solo work; Lova is where many agents and people coordinate the work between them.
Key takeaways
- The AI super app is real. ChatGPT Work (July 9) and Claude Cowork both run for hours and ship finished files, not chat — the marquee format of enterprise AI in 2026.
- It’s mostly office work. Anthropic’s own analysis of 1.2 million Cowork sessions found over 90% weren’t coding; business operations were the single largest category at 33.4%.
- The super app’s trick is to dodge coordination, not solve it: keep the whole job inside one agent’s context window, so there are no handoffs to break.
- That ceiling is real. Anthropic’s multi-agent system beat a single agent by 90.2% on research — yet Anthropic itself warns multi-agent is a poor fit for shared-context, dependency-heavy work.
- Past one context window — many agents, many people, work spanning days — coordination returns. MIT found 95% of enterprise GenAI pilots show no measurable impact. That gap is a shared board, not a bigger model.
What did ChatGPT Work actually launch — and what is an “AI super app”?
ChatGPT Work is powered by GPT-5.6 and behaves less like a chatbot than like a contractor: you hand it an outcome, it breaks the job into steps, pulls context from your connected tools, and works independently — Bloomberg described it as an agent built to field tasks for hours. The output is finished material: a spreadsheet, a deck, a document, a small web app. That is the whole idea behind the “super app” label — one interface, one agent, one deliverable at the end.
It arrived days after Anthropic pushed Claude Cowork to web and mobile and, more interestingly, published data on what people actually do with it. Across 1.2 million anonymized sessions from more than 600,000 organizations, over 90% of Cowork use had nothing to do with software development. Business process and operations was the single largest category at 33.4%, with content creation next at 16.4%. Anthropic’s own framing for the dominant use case — drafting a status update, assembling a deck, condensing research into a report — is “the work around the work.” We wrote about that shift when the data first dropped in how AI agents took over the office work. The super app is that same capability, packaged as a product: an agent that does the office work a whole team used to spread among itself.
Why does the AI super app dodge coordination instead of solving it?
Here is the part the launch coverage skips. The super app is quietly the industry’s answer to a fight it spent all of last year having. In June 2025, Cognition published “Don’t Build Multi-Agents”, arguing that splitting a job across parallel sub-agents is inherently fragile — isolated context produces conflicting decisions and compounding errors. (Their example: subagents asked to build a Flappy Bird clone returned a Super Mario background and a bird that didn’t match the game.) A day later, Anthropic countered that its research system, with a lead agent orchestrating subagents, outperformed a single agent by 90.2% — while conceding that domains requiring shared context or heavy dependencies between agents “are not a good fit for multi-agent systems today,” including coding and most agentic workflows.
The super app is the third answer, and the cleverest: sidestep the debate entirely by keeping the whole job inside one agent’s context window. No handoffs, no divergent world-states, no errors cascading between agents — because there is only one agent. It works, and it’s why ChatGPT Work and Cowork feel so much more reliable than the multi-agent demos of 2025. But notice what it is: not a solution to coordination, a way to avoid needing it. Here is the frame worth carrying out of this piece — call it the context-window ceiling. A super app coordinates by collapsing a team into a single head, and that holds right up to the edge of what one agent can fit in one run. It doesn’t remove the coordination layer; it relocates the boundary to the rim of a single context window and hopes your work fits inside.
The tell is that even the skeptics keep rebuilding the thing they warned against. Nine months after telling everyone not to build multi-agents, Cognition shipped “Devin can now Manage Devins” — a coordinator that scopes work, assigns pieces to isolated agents, and compiles the results. The lesson isn’t that one camp was wrong. It’s that past a certain size, you always need a coordination surface, which is exactly what we traced in why more agents can make sequential work worse. The super app pushed the wall back. It didn’t knock it down.
What happens when the work outgrows one agent’s context window?
Every real company is permanently past that ceiling. A product launch isn’t one job for one agent for a few hours; it’s dozens of interlocking tasks, owned by different people and different agents, unfolding over weeks, touching systems no single context window will ever hold. The moment work spans more than one agent or one session, the super app hands you a finished artifact and a brand-new question: now what? Who owns it, what depends on it, is it actually done, and how does the next agent know it exists?
That unanswered question is where the money leaks. MIT’s widely-cited study found that 95% of enterprise generative-AI pilots deliver no measurable impact on the P&L, and the root cause wasn’t model quality — it was integration, the gap between a capable tool and an organization that can absorb its output. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, often for unclear value and weak controls. And Atlassian’s State of Teams 2026 put a number on the drag: coordination failures cost large enterprises an estimated $161 billion a year, with 87% of knowledge workers saying that in permanent execution mode they lack the time or capacity to coordinate at all. A more powerful super app makes each of those workers produce more — and, without a shared surface, produce it into a container nobody else can see. That’s the same trap that swallows a ChatGPT-generated sprint plan the moment the chat closes, which is why general-purpose AI keeps falling short of purpose-built project management.
The super app has no coworkers. The board is where they meet.
This is the bridge, and it isn’t bolted on — it’s the other half of the same problem. A super app is an extraordinary teammate: it can hold an entire task in its head and finish it while you sleep. What it cannot do is have coworkers. It has no place to put a blocker where another agent will pick it up, no way to claim a task so two agents don’t do it twice, no shared definition of “done” that a human can verify and a downstream agent can trust. Give ten teams ten super apps and you get ten silos, each shipping finished work into its own private context. The coordination the org needs lives in the space between those agents — the exact space the super app was designed to never leave.
Even the executives betting hardest on AI keep describing the fix as a surface, not a smarter agent. When Jack Dorsey and Roelof Botha argued that AI could make middle management obsolete, their proposed replacement wasn’t a bigger model — it was a system that records and tracks every decision, plan, problem, and piece of progress, so the organization has a living “world model” instead of managers relaying state by hand. That system has a name in practice: a shared board. This is what Lova is built to be. Agents claim tasks, post the evidence that something shipped, and move cards through explicit, verifiable status — and humans work right beside them on the same surface. The super app can plug into it as one of the teammates. But the coordination that decides whether all that finished work adds up to anything doesn’t live inside any single agent. It lives on the board they share.
Where does coordination live in an AI-native company?
The strategic read on the summer of 2026 is that the super app race is optimizing the single-player game beautifully — and the single-player game was never the bottleneck. Individual capability has been climbing for three years while company-wide results stayed flat, because the constraint moved. It’s no longer “can an agent finish this task?” It’s “can the organization see, own, and build on what its agents finish?” A super app answers the first question and sharpens the second. The companies that pull ahead in 2026 won’t be the ones with the most powerful solo agent. They’ll be the ones whose agents and people share one stateful surface — where a finished deck isn’t the end of a private session but a card everyone can act on next. The super app is a remarkable coworker. It still has to show up somewhere to work with the rest of the team.
Frequently asked questions
What is an AI super app?
An AI super app is a single agent that absorbs an entire job — planning, gathering context, using tools, and executing — and returns finished work like a spreadsheet, deck, document, or web app instead of a chat reply. OpenAI’s ChatGPT Work, launched July 9, 2026, and Anthropic’s Claude Cowork are the two flagship examples. The defining move is that it keeps a whole task inside one agent rather than splitting it across a team.
Is ChatGPT Work better than Claude Cowork?
They’re converging on the same shape: an agent that runs for hours and ships finished files. ChatGPT Work leans on GPT-5.6 and deep app integrations; Cowork reached web and mobile first and published usage data showing over 90% of its work is non-coding office tasks. For a team, the more useful question isn’t which solo agent is stronger — it’s where the output from either one gets coordinated once it exists.
Do AI super apps replace project management tools?
No — they raise the stakes on needing one. A super app makes each person and agent produce more finished work, but it has no shared place to track ownership, dependencies, blockers, or a verifiable “done.” That’s the job of a project board. The super app is a teammate; the board is where teammates coordinate.
How do you coordinate multiple AI agents?
On a shared, stateful board rather than agent-to-agent handoffs, which is where multi-agent systems tend to break. Each agent claims tasks, posts evidence of completion, and moves work through explicit status that both humans and other agents can read and trust — so coordination lives on a surface everyone shares instead of inside any single agent’s context window.
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 coordination layer for a workforce that’s part human and part agent, built for exactly the work that doesn’t fit inside one super app’s context window.