Taste is not the err in human input

Agents and swarms of agents are quite nifty nowdays. My current setup is simple. It runs a coordinator that I communicate with, and that delegates work to a others. The flow is generally: "Hey, I need this feature from roadmap. Ask questions if needed, if I don't answer in due time, make the best call with our north star." -> Coordinator makes a plan for dev/qa -> dev devs -> pushes to branch when tests are green -> qa verifies results on dev server -> loopy-loop until obvious bugs are fixed. Review bot makes review for PR. Later on I test the feat to see if it's anything like I wanted and correct, then merge when good and pool to releases that are gated. Currently working on trying to get the agents to know what is good in my opinion.

I think the current killer combo is the ability to be creative and have some propriatory data, as that's something that can't be accessed by competition. The cost of iteration is low. So low that I see no point in using a wireframe demo for the ui, as simple paper sketches can be read by AI to close enough for live iteration for actually working software. And most rigid general UI's are going away too in favor of agents generating a personalized UX on the fly, so this is just temporary. Omarchy is the beacon for this, a machine that is ever evolving to your needs.

The stuff that stays valuable is hard to automate - having taste, accountability, human relationships and even defining what is good. No one knows what the future will be, but it will be awesome.

Kalle Tolonen
22.9.2026

The human input

We can now think big. In early 2026 I was chatting at work that a greenfield project that needed 10 devs, would be 1 dev by year end. It's not year end and I'm pretty sure I could handle the dev load of similar proportions, given enough access to tokens, naturally.

Software is being eaten up by inference. Why have rigid structures, if they can bend instead? We need some structure for serving and transforming data, combining private and public data for now, but once inference becomes cheap enough, I think that moat of rigidness will succumb too. I even have positive experiences in data-analysis that's fast to do with agents, so on the fly generated dynamic dashboards only need a good research question.

This particular case would also benefit from gis-analysis, as the distances in rural Finland are larger, so the costs for service providers are also bigger. So one more round of polish needed to gain valuable insight on where the potential customers are.

There is still quite a lot of taste needed for course correction and agile is the right way to go, and we're in the endless march of 9's for producing quality sw. I was perfectly happy with earlier models, that is something you need to remember - you will always lust for more.

Process thoughts

We still need a good hunch on what we're aiming for in the big picture:

• how many users and who they are
• what security measures (derived from above points - internal sw used by 3 people with only GET-api's is different than a public bank)
• what tools are allowed
• general design vibe
• what platforms (mobile/pwa, native apps, web, desktop ui/cli)
• now we can decide tech platforms (security and use case dictate this)

Then we build the env's needed, those are dictated by the needs of the project. Less is more, and we should opt for cloud computing in every step of the way to keep cadence up, since the days of "my env is broken" are really fucking expensive, when you're doing work that used to take 10 people a year ago, even as the cost curve is coming down. Comp set isn't sitting still worrying about your docker hell. On the other hand - agents are really good at configuring too now days - just try Omarchy.

Skills needed for sw dev in 2026

Not in any particular order.

• People skills
• Project management skills, steering with right-sized inputs (not token-wise, but concept-wise)
• a level of tech understanding, so that you can steer a team/horde of ultra capable robots (or better, know how to outsource that)
• Creativity
• Having a hunch of what the agents can and can't do (graphics still mostly suck)

Details matter, but in a different way than before

There used to be a day and age, when writing code that was eloquant, efficient and beautiful was really meaningful. I've written some, but honestly most of it was not that spectacular. And that's true for the most of the code I've reviewed. People who have worked at legacy projects know this.

You should get some people skills, as everything done by you, can be done by a computer, and if it can be done by a computer, it can be automated. Only enough chaos and noise brings (digital) job safety. But you can't outsource the taste still. What we are building and why.

it's increasingly clear that you need to keep running at an accelerating phase and find your place where you can produce some value that someone is willing to pay for to earn you keep. That has been the story of humanity, and the pace is accelerating. Last year I thought that AI is pretty good and this year the harnesses and running the agents on their computer have made the switch to amazing for me.