Beyond the editor
Not long ago, AI coding tools lived in the editor, finishing a line or drafting a function. Now they work across the whole project. In a single week, agents can summarize research, turn a kickoff call into a plan, sketch interface options, write and refactor production code, and check the result against what was agreed.
That changes what an agent is for. It is less a faster keyboard and more a collaborator that reads the context, proposes a change and does the legwork in between. The question stops being whether it can write the code and becomes what it should be pointed at.
The speed is real
The productivity gains are not marketing. Work that used to fill a sprint now takes days, and some of what took days now takes an afternoon. A prototype good enough to put in front of a client can exist before the meeting ends. The refactor nobody wanted to schedule simply gets done.
The gains are largest in wide, repetitive, easy-to-check work: migrating a codebase, rebuilding templates, checking every page after a change. That is exactly the work agents now carry, and it frees people for the parts of a project that need them.
Workflow is the work
The same tools produce very different results in different hands, and the difference is rarely the model. It is the workflow around it: how much context the agent is given, how the task is broken down, and how the output is checked.
Ours is deliberately plain. Give the agent the real constraints: the brand rules, the type scale, the spacing system, the things that must never change. Ask for one change at a time. Look at it in the browser, at every screen size, before moving on. Small steps with a review after each beat large leaps that need untangling later.
The loop is the method
Agents make iteration cheap, and cheap iteration changes how a project feels. Instead of debating a layout in the abstract, we build three versions and look at them. Instead of describing an animation, we try it, watch it, and adjust.
A single screen can go through a dozen rounds in an afternoon: move the title, tighten the spacing, try the image at a different crop, look again. Each round takes minutes, not days.
“Make, look, refine. The cycle is now the fastest path to good work, as long as someone is looking carefully at every turn.”
Taste is the bottleneck
When anything can be generated in minutes, the scarce thing is knowing which version is right. Agents produce plausible work at volume. Plausible is not the same as good, and it is a long way from distinctive.
The hard calls are rarely about whether something can be built. On QuantLink, the product had plenty of good features; the work was deciding what people see first, where they start and what to leave out. An agent can draft every option. Choosing the one that fits, and noticing what feels slightly off, is still taste.
Judgement doesn't delegate
Taste decides what is good. Judgement decides what is ready to ship, and that doesn't outsource well either. An agent will write code that passes its own tests and still misses the point of the feature. It will make the change you asked for without knowing the change was a mistake.
We built Sands, our own agent for project work, around that limit. It proposes; people decide. The same rule holds for any team using these tools: the agent does the work, and a person who understands the project owns the decision.
Where it goes next
These tools are changing month to month. The agents we use now handle longer tasks, hold more context and recover from their own mistakes better than the ones we started the year with. The scale is changing too: one person can work with a single agent, direct a lead agent that manages ten, or run a hundred in parallel. The pace shows no sign of slowing.
The gap between teams will widen, but not along the line of who uses AI. Everyone will. The difference will be who pairs these tools with a clear point of view: a workflow built for review, a loop tight enough to iterate quickly, and people with the eye to know when the work is right. The tools make good teams faster. They don't make taste.
What's worth making
Put together, this means a small team can now cover ground that once took a much larger one. Research, design, code and the rounds of iteration between them move faster than they ever have, and they will keep getting faster.
But speed was never the hard part. Anyone can ask an agent for twenty versions of a page. Someone still has to look at all twenty, know which one is right, and know why. Agents make the work cheap to produce. They leave the deciding to us.
That is how we try to work at Espyus: a small team that uses these tools every day and spends the time they save on the decisions that matter. Let the agents do the making. Keep the judgement close.
If you are working out where agents fit in your own team, or you have something worth making, we would like to hear about it.


