The tool does 80% of the work; knowing which one is the skill. That is the sentence I kept coming back to this month, watching two announcements that looked unrelated and are actually about the same thing. On August 25, a major Chinese internet company released its agent product line for productivity scenarios, deeply integrated with its own collaboration suite. On the same day, a leading American automaker began pushing a large language model into the infotainment systems of its China-market cars. Different industries, same direction: the software now carries the routine load, and the human decides which load goes where.
Here is the plain version, the way I would explain it across a workbench. An agent, in this context, is software that does not just answer when you ask — it carries out a task you delegate: gathering the documents, drafting the report, sending the message, end to end. The car update is the same idea in a different room: the voice assistant understands a fuller request and handles more of the steps, while the core driving system stays in the company’s own hands.
Why the word agent matters more than the products
Take the word apart first. Assistant meant you talked and it talked back. Agent means you delegate and it delivers. That shift in the verb is the whole story. The productivity launch is built around this: you describe the job, and the software moves through the steps, pulling from the collaboration suite, checking the materials, producing the result. It is the difference between a tool you operate and a tool that operates on your behalf.
The car side is the same shift wearing a dashboard. Instead of a voice system that recognises a command and searches, you get one that holds context and completes a chain — navigate, book, message, in one flow. The driving itself is not part of the deal; the announcement was explicit that the autonomous core remains proprietary. The model is in the car to make the interaction smarter, not to take the wheel.
That boundary is worth noticing, because it is the correct engineering instinct. Put the AI where it raises the experience, keep the safety-critical system where you need full control. That is exactly how a careful builder draws the line — the tool handles what is safe to delegate, and the critical path stays in expert hands.
What the 80 percent rule means in practice
Here is where the workbench wisdom applies directly. A tool that does 80 percent of the work is only useful if you know what the 80 percent is. In a kitchen, a good knife does most of the chopping, but you still decide what to chop, how fine, and what to do with it after. The new agents are the same: they will carry the routine, structured, repeated tasks — the reports, the summaries, the message flows — and they will do it fast.
The remaining 20 percent is the part that does not transfer. Judgement about what matters, decisions about what to include and exclude, the call on tone and intent, the moment when something looks wrong and someone has to stop and say so. No agent announcement has claimed that part, and no serious one should. The tool is powerful; the person is still the one with the standard.
I had to correct my first reaction to these launches, because my instinct was to frame them as either a marvel or a threat. No, that is not quite right. The more accurate read is that they are a milestone in delegation — software reaching the point where it can take on whole routine jobs, not just fragments. Whether that is good or bad news depends entirely on what you choose to do with the freed time, and that is not a question the tool can answer.
How to treat an agent like a new employee
For anyone who wants to actually use this, here’s how I would start. Treat the agent the way you would treat a capable new hire on day one: do not hand it the crown jewels, and do not leave it idle. Give it one small, well-defined, repeatable task. Watch how it handles the boundaries. Check its work before you trust it. Then hand it the next task, slightly larger.
The analogy holds because the failure modes are the same. A new employee who is given vague instructions will produce vague results. An agent given a fuzzy description will produce fuzzy output that looks confident. The skill is in the brief: the clearer you are about the goal, the format, the constraints, the better the result. That is not a software skill. It is the oldest management skill there is, now applied to a tool that does not need a desk.
You’ll see the difference quickly between someone who treats these tools as toys — asking for poems and jokes, impressed for a week — and someone who treats them as staff. The first is entertainment. The second compounds. Over a few months, the difference in output is not a matter of taste; it is a matter of who used the tool as a tool and who used it as a spectacle.
The concrete scene: a Friday afternoon
Picture the Friday afternoon this is aiming at. The weekly report needs assembling from scattered documents, the meeting notes need sending to six people, the follow-up messages need drafting. In the old routine, that is an hour of clicking, copying and rephrasing. With an agent holding the context of your collaboration tools, it is one clear instruction and a review pass.
That hour is the 80 percent. What do you do with it? The answer is the entire point of the tool, and it is the part no product launch can settle for you. Whether the saved hour goes into the report that actually gets read, the planning that moves the project, or the evening that keeps you sane — that is the 20 percent that decides whether the tool makes your work better or merely faster.
The feel of it, if you have ever handed off a task to a reliable colleague, is familiar: a slight hesitation the first time, then a steady glide once you see the standard hold. The same trust curve applies. The first delegation is the hard one; every delegation after it gets easier as you calibrate what to hand over and what to keep.
The honest limits, stated without hype
Let me be straight about the boundaries, because the launches are polished and the reality is rougher. Agents will make mistakes, especially on tasks that involve judgement disguised as routine. They will occasionally do the wrong thing with confidence, which is the most dangerous failure mode a tool can have. And the integration depth varies — being inside one ecosystem is not the same as being useful everywhere.
There is also the question of what gets automated into brittleness. A workflow that runs itself is wonderful until something changes and the tool keeps doing the old thing faithfully. That is why the human review pass matters, and why the 20 percent that stays with the person is not a leftover; it is the safeguard. The tool that runs unsupervised is a tool that will eventually embarrass its owner.
What can be said with confidence is that the direction is clear. Software has crossed from answering to doing, in the office and in the car, and the boundary the automaker drew — smart interface, human-held critical control — is the boundary every careful user should draw too. Delegate the routine, keep the judgement, check the work.
The verdict from the workbench
So take this month’s two announcements apart and the parts are these: a productivity agent that carries routine office work end to end, and a car assistant that holds context across a whole request while the driving stays firmly human-owned. Different rooms, same architecture: delegate the load, keep the decisions.
The feel of it, ultimately, is not futuristic at all. It is the old workbench lesson arriving in new packaging — the tool does 80 percent of the work, and the skill is knowing which 80 percent, and which 20 percent to keep for yourself. Hands-on beats theory every time, and the hands-on test here is simple: give the tool one real task this week, watch it work, and decide where the line belongs for you. The tools will keep improving either way. The only question that is yours to answer is what you do with the part that does not transfer.
The deeper shift behind both announcements
Step back from the products and the shape of the change becomes clear. Both launches are built on the same architectural idea: the model holds context, and the software around it executes steps. That combination — understanding plus action, wired into an existing ecosystem — is what separates an agent from a chatbot. A chatbot is a conversation. An agent is a workflow with the human moved to the edges, setting the goals and reviewing the output.
This matters because it changes what proficiency means. Proficiency used to mean knowing how to operate the software — which menu, which shortcut, which sequence. With agents, that kind of knowledge shrinks in value, because the tool carries the sequence itself. What grows in value is the ability to define the task well: to know what good looks like, to specify it clearly, and to recognise when the result is wrong. That is a different skill set, and it is available to people who have never touched a shortcut key.
For the person who wants to stay ahead of the curve, the practical implication is worth writing down. Do not spend this year memorising features; spend it practising specification — writing clear briefs, setting standards, checking outputs. That practice survives every tool upgrade, because it is the skill of being the person who decides, and the person who decides is the person who keeps the 20 percent. Here’s how you’ll see it play out: the people who thrive in the agent era are not the ones who know the tool best. They are the ones who know the work best, and can say exactly what they want done.
One final note, because I want the takeaway to be practical rather than philosophical. This month’s launches do not require you to change your habits overnight, and anyone who tells you otherwise is selling urgency. What they do is lower the cost of trying: the productivity agent sits inside tools many people already use, and the car assistant arrives as an update rather than a new purchase. That means the testing ground is already around you. Pick one real task, hand it over, and watch. The 80 percent will show you where it lives, and the 20 percent will show you where you live. That is the whole lesson, and it fits on a workbench.